A Multimodal Framework for Prognostic Modelling of Mental Health Treatment and Recovery Trajectories

Abstract

The clinical management of major depressive disorder is constrained by a trial-and-error approach. The clinical management of major depressive disorder is constrained by a trial-and-error approach. While computational methods have focused on static binary classification (e.g., responder vs. non-responder), they ignore the dynamic nature of recovery. Building upon the recently proposed prognostic theory of treatment response, this article presents a methodological framework for its operationalisation. We define a multi-modal data architecture for the theory’s core constructs—the Patient State Vector (PSV), Therapeutic Impulse Function (TIF), and Predicted Recovery Trajectory (PRT)—transforming them from abstract concepts into specified computational inputs. To model the asynchronous interactions between these components, we specify a Time-Aware Long Short-Term Memory (LSTM) architecture, providing explicit mathematical formulations for time-decay gates to handle irregular clinical sampling. Furthermore, we outline a synthetic validation protocol to benchmark this dynamic approach against static baselines. By integrating these technical specifications with a translational pipeline for Explainable AI (XAI) and ethical governance, this paper provides the necessary blueprint to transition psychiatry from theoretical prognosis to empirical forecasting.

1. Introduction

1.1. The Prognostic Imperative in Psychiatry

The management of major depressive disorder (MDD) remains one of the most significant challenges in contemporary medicine. Despite a wide array of available treatments, the process of finding an effective intervention for an individual patient is often a protracted and painful exercise in trial and error [1,2,3,4,5]. Response rates to initial pharmacotherapy are modest, with fewer than half of patients experiencing a clinically meaningful response and only a third achieving full remission [6]. This variability underscores the limitations of current clinical decision-making approaches, which rely heavily on sequential empirical adjustments rather than predictive guidance [7,8]. Such formulations fail to capture the continuous and context-dependent nature of recovery, reducing a multidimensional temporal process to a static label.

For decades, the field of personalised psychiatry has pursued reliable biomarkers to guide treatment selection, yet this goal has remained elusive [9]. In response, a substantial body of research has applied machine learning and other computational techniques to predict treatment outcomes [5,10,11,12,13]. However, these efforts have had minimal impact on clinical practice, a translational failure that stems from the dominance of static prediction models and their limited integration into clinical workflows: the dominance of the static prediction paradigm [14]. The vast majority of current models are designed to classify a single, binary endpoint—such as “responder” versus “non-responder”—at a fixed future time point, based on a snapshot of baseline characteristics [15,16]. This approach is a profound oversimplification of clinical reality. Recovery is not a singular event but a dynamic, longitudinal process with significant intra-individual variability [17,18]. By collapsing this rich temporal information into a binary outcome, static models discard crucial data about the speed of response, patterns of symptom fluctuation, and the stability of improvement, all of which are vital for ongoing clinical management [19,20,21].

The critical knowledge gap, therefore, is not merely in improving the accuracy of static prediction, but in developing the capacity for dynamic, personalised prognosis—the ability to forecast the trajectory of a patient’s illness course under a specific treatment [22]. A clinician’s task is not just to guess an outcome at 12 weeks but to monitor and adjust treatment based on the patient’s progress over time. A truly useful clinical tool must align with this reality. It must forecast the path, not merely the endpoint. This shift from a static to a dynamic perspective represents the central computational and clinical imperative for the next generation of personalised psychiatry.

1.2. A Recap of the Prognostic Theory

To address this imperative, a prior work introduced a theoretical framework conceptualising mental health treatment response as a complex dynamic system [23]. This framework establishes an interdisciplinary paradigm by drawing a powerful analogy from the engineering field of structural health monitoring and its advanced application, damage prognosis [24]. While diagnosis in engineering answers the question, “What is the state of the system now?”, prognosis answers the far more valuable question, “Given the current state and future loads, how will the system behave over time?” [24]. The goal is to forecast the “remaining useful life” of a system, enabling predictive and adaptive interventions.

Applying this prognostic perspective to psychiatry, an individual patient is conceptualised as a complex system, with MDD representing an undesirable but stable “attractor state”. A therapeutic intervention is an active perturbation or “load” applied to this system with the goal of shifting it toward a stable state of recovery [25,26]. To formalise this, the theory introduced three core constructs [23]:

• The Patient State Vector (PSV): A comprehensive, multi-modal, and time-varying representation of a patient’s state, analogous to the initial state assessment in engineering. It integrates clinical, biological, and high-frequency digital phenotype data to create a high-dimensional characterisation of the individual.
• The Therapeutic Impulse Function (TIF): A formal characterisation of a treatment’s properties (e.g., pharmacodynamics, therapeutic modality), analogous to the “future loading conditions” in an engineering system. It defines the specific perturbation being applied to the patient’s system.
• The Predicted Recovery Trajectory (PRT): The forecasted, continuous path of a patient’s symptom severity over time, analogous to the “remaining useful life” prediction. The theory’s central thesis is that the PRT is an emergent property of the dynamic interaction between an individual’s unique PSV and a specific TIF.

Together, these constructs provide a formal vocabulary for linking patient data, treatment properties, and temporal outcomes within a unified prognostic modelling framework.

1.3. Objective: From Theory to a Testable Methodology

The formulation of this prognostic theory was the necessary first step. However, a theory’s value is ultimately determined by its ability to be tested, refined, and translated into practice. This methodological contribution bridges that gap between theoretical formulation and practical implementation by presenting a structured methodological framework that operationalises the prognostic theory, making it empirically testable and directly applicable to computational and clinical research contexts.

Rather than reporting empirical findings, this work provides a replicable and adaptable methodology for researchers aiming to apply and validate the prognostic theory. It responds to persistent methodological barriers in computational psychiatry—heterogeneous data standards, limited sample sizes, and insufficient validation protocols—that have constrained translational progress [27,28,29]. By establishing a standardised, transparent, and ethically grounded process, this framework aims to serve as a shared foundation for reproducible research and for the design of human-centred clinical support tools.

The paper is structured as follows: Section 2 provides a granular guide to operationalising the core constructs, detailing the data architecture for the PSV, TIF, and PRT. Section 3 proposes and justifies a specific deep learning architecture—a time-aware Long Short-Term Memory (LSTM) network—for forecasting the PRT and outlines a rigorous protocol for its training and evaluation. Section 4 outlines the translational pathway, including explainability, ethical governance, and principles of human-computer interaction for clinical integration. Section 5 discusses the implications and limitations of the framework, positioning it as a scalable foundation for predictive, personalised, and human-centred psychiatry.

This framework situates the development of prognostic AI within a human-centred design paradigm, emphasising interpretability, ethical responsibility, and the co-evolution of computational and clinical reasoning.

1.4. Scope and Contributions

This manuscript presents a methodological framework—a theoretical and architectural blueprint—rather than the results of an empirical trial. Given the current lack of standardised, multi-modal datasets combining high-frequency digital phenotyping with biological markers, a rigorous “design-first” approach is required before large-scale data collection can commence.

The specific contributions of this work are:

• A Novel Taxonomy for Prognosis: The operationalisation of the “Patient State Vector” (PSV) and “Therapeutic Impulse Function” (TIF) provides a formal vocabulary for modelling the dynamic interaction between patient and treatment, distinct from static baseline predictors.
• A Time-Aware Architecture Specification: We specify a deep learning architecture explicitly designed for the irregularity of real-world clinical data, addressing the “asynchrony” problem that renders standard RNNs ineffective in practice.
• A Translational Roadmap: Unlike purely technical papers, we provide an integrated implementation strategy that treats Explainable AI (XAI) and ethical governance as architectural requirements, not optional add-ons.

Limitations of Scope: This paper does not report results from a clinical cohort. Instead, it defines the protocol and computational specifications required to generate such results. Validation strategies, including synthetic data simulation, are outlined in Section 3.3 to guide future empirical work.

2. From Theory to Data: Operationalising the Prognostic Constructs

A prognostic model is only as powerful as the data it learns from. Building on the theoretical constructs introduced earlier—the Patient State Vector (PSV), Therapeutic Impulse Function (TIF), and Predicted Recovery Trajectory (PRT)—this section describes how these components are operationalised as multi-modal data structures suitable for computational modelling and empirical validation. This section provides a detailed blueprint for this operationalisation, specifying the data sources, feature engineering pipelines, and encoding strategies that form the inputs and outputs of the proposed prognostic model. The emphasis is placed on creating a data architecture that is both technically robust and compatible with clinical and ethical standards for real-world deployment.

2.1. The Patient State Vector (PSV): A Multi-Modal, High-Frequency Data Architecture

The PSV represents a comprehensive, temporally dynamic description of a patient’s state [23]. Its construction requires the integration of heterogeneous data streams, each with different temporal resolutions and characteristics. A key innovation of the PSV is its departure from a simple, flat baseline feature vector. Unlike traditional baseline feature sets, the PSV is designed as a chronologically aware, multi-scale representation of an evolving system. It combines static features (e.g., genetics), low-frequency time-series (e.g., weekly clinical assessments), and high-frequency time-series (e.g., continuous sensor data). This nested temporal structure introduces a significant methodological challenge: the model must not only fuse heterogeneous data types but also capture the complex cross-scale dependencies that occur over time. Addressing this requires data fusion strategies that preserve temporal alignment and contextual relationships, going beyond simple feature concatenation. The following subsections outline the four key data domains and the process through which they are integrated into the PSV architecture.

2.1.1. Clinical Data

This domain includes the information typically collected in clinical trials and practice. While often considered “baseline” data, a dynamic prognostic approach requires capturing these measures longitudinally.

• Data Sources: Electronic Health Records (EHRs), clinical interviews, and Patient-Reported Outcome Measures (PROMs).
• Metrics:

  –

  Symptom Severity: Standardised scales such as the Patient Health Questionnaire-9 (PHQ-9) and the Hamilton Depression Rating Scale (HAM-D), collected at regular intervals (e.g., weekly) to track the evolution of symptoms.

  –

  Diagnostic History: Coded diagnoses (e.g., ICD-10) for MDD, comorbidities (e.g., anxiety disorders, substance use disorders), and history of prior treatment attempts.

  –

  Demographics: Age, gender, socioeconomic status, and other relevant demographic variables.

• Feature Engineering: Raw scores from clinical scales are used directly. Diagnostic history can be one-hot encoded. Longitudinal scores form a low-frequency time-series input to the model. Data harmonisation and standardisation protocols are essential to ensure comparability across sources and enable reproducible research.

2.1.2. Biological Data

This domain captures the neurobiological and physiological substrates that may influence or moderate treatment response. These data sources tend to be sparse, expensive to obtain, and typically static or low-frequency in nature, yet they provide essential context for individual variability.

• Data Sources: Genetic assays, neuroimaging scans, and blood or saliva samples.
• Metrics:

  –

  Pharmacogenetics: Genotypes for key genes influencing drug metabolism (e.g., cytochrome P450 enzymes like CYP2D6, CYP2C19) or drug targets [30].

  –

  Neuroimaging: Measures derived from resting-state functional MRI (fMRI), such as functional connectivity within and between key networks (e.g., default mode network, salience network), or structural measures from T1-weighted MRI [31,32,33,34,35].

  –

  Peripheral Biomarkers: Levels of inflammatory markers (e.g., high-sensitivity C-reactive protein), stress hormones (e.g., cortisol), and neurotrophic factors [36,37,38].

• Feature Engineering: Genetic polymorphisms are encoded as categorical variables; neuroimaging and biomarker data are represented as numerical features. Because these data are static or extremely low-frequency, they function as conditioning variables that modulate the model’s dynamic predictions rather than forming a temporal sequence. Clear documentation of acquisition protocols and feature preprocessing is necessary to ensure replicability and comparability across datasets [39].

2.1.3. Digital Phenotype Data

This domain is a cornerstone of the PSV, providing a continuous, objective, and ecologically valid view of real-world behaviour [40]. It involves transforming high-frequency sensor data from personal devices into clinically meaningful behavioural indicators [40,41]. From an HCI perspective, these data also reflect the interaction between individuals and their digital environments, making them central to human-centred prognostic design.

• Data Sources: Passively collected data from sensors embedded in smartphones and wearables (e.g., GPS, accelerometer, gyroscope, screen usage, call/text logs) [40,41].
• Feature Engineering Pipeline: Raw sensor data is typically aggregated into daily or hourly summaries to align with the temporal scale of mood fluctuations. Major behavioural domains and representative features include:

  –

  Mobility Patterns: Raw GPS coordinates are processed to derive features reflecting behavioural activation and social withdrawal. These include:

  • Location Variance: The statistical variance of latitude and longitude, capturing the geographic spread of a person’s movement.
  • Entropy: A measure of the diversity and predictability of visited locations. Lower entropy suggests a more restricted and repetitive routine.
  • Time Spent at Home: The proportion of time a user’s device is located at their inferred home location.
  • Distance Travelled: Total daily distance covered.

  –

  Social Activity: Call and SMS text message logs (metadata only, not content) are used to quantify social engagement. Features include:

  • Number and Duration of Incoming/Outgoing Calls: A proxy for active social interaction.
  • Number of Unique Contacts: A measure of social network size.
  • Text Message Frequency: A proxy for passive social communication.

  –

  Circadian Rhythms: The regularity of daily routines, particularly sleep-wake cycles, is a critical indicator of mental health. These patterns can be inferred from multiple sensors:

  • Phone Lock/Unlock Patterns: The longest continuous period of no screen interaction during nighttime hours can serve as a proxy for sleep duration [40,41].
  • Accelerometer Data: Periods of prolonged inactivity from a wearable can more directly measure sleep.
  • Circadian Regularity: A metric quantifying the stability of these patterns from day to day.

  –

  Physical Activity: Accelerometer data is used to derive standard activity metrics like daily step count and time spent in different activity intensity levels (e.g., sedentary, light, moderate).

Digital phenotype features are high-frequency inputs that provide temporal granularity and behavioural context. When integrated with clinical and biological data, they enhance both model accuracy and interpretability, particularly within user-facing systems that visualise behavioural trajectories for clinical feedback.

Smartphone-based sensing is subject to sampling variability across devices and operating systems. GPS sampling may fluctuate depending on device movement, power-saving modes, or environmental conditions. Accelerometer sampling rates may degrade under battery optimisation settings or background execution limits. These sources of variability require normalisation procedures to ensure consistent daily feature extraction.

Noise and artefact removal is performed prior to aggregation: implausible GPS jumps are removed using speed thresholds, stationary accelerometer noise is filtered using standard deviation-based heuristics, and missingness arising from battery depletion or user behaviour is explicitly encoded. These corrections ensure that the temporal patterns captured in the PSV reflect behavioural signals rather than hardware or OS artefacts.

Although this framework focuses primarily on smartphone-based sensing, it is fully compatible with wearable accelerometers and physiological sensors such as heart rate variability (HRV) or electrodermal activity (EDA). These additional streams can be integrated into the PSV as higher-frequency inputs for enhanced temporal resolution.

Sensor Data Acquisition Pipeline

The digital phenotype features used in this framework originate from continuous streams collected through smartphones and wearables. GPS coordinates are typically sampled at 1 Hz, accelerometer data at 50–100 Hz depending on device settings, and screen interaction events are captured as discrete timestamps. All raw timestamps are first converted into a unified time base (UTC), and device-specific clock drift is corrected when available.

To enable integration within the PSV, raw high-frequency streams are aggregated into daily summaries through a preprocessing pipeline consisting of: (1) timestamp alignment, (2) noise and artefact removal (e.g., invalid GPS points, accelerometer spikes), (3) interpolation for short gaps, and (4) extraction of behavioural indicators such as location entropy, inferred sleep duration, circadian regularity, and mobility radius.

Longer gaps due to battery loss, sensor deactivation, or device shutdown are handled by explicit missingness markers, ensuring that the time-aware LSTM can model irregular sampling intervals [42]. This pipeline provides a structured and reproducible method for converting heterogeneous sensor data into the dynamic features of the PSV.

2.1.4. Data Integration and Handling

The PSV constitutes an inherently multi-modal, multi-rate, and partially asynchronous time-series dataset, posing challenges in data fusion, missingness, and temporal alignment [43,44]. A reproducible methodological framework must address these challenges through well-defined preprocessing and harmonisation procedures.

• Aggregation: High-frequency digital phenotype data is aggregated to a consistent temporal resolution (e.g., daily summaries).
• Imputation: For sparse data streams like weekly clinical scores or occasional biomarkers, appropriate imputation or interpolation techniques (e.g., mean imputation, forward-fill, or more sophisticated model-based imputation) must be applied.
• Asynchrony: The asynchronous nature of data collection across modalities is a core temporal modelling challenge. This can be mitigated at both the preprocessing and modelling stages. Architecturally, time-aware mechanisms, such as the gated structures described in Section 3, can explicitly represent time intervals between observations and learn from irregular sampling patterns [45,46,47].

To support implementation,

Table 1. The Patient State Vector (PSV) Data Dictionary.

Domain	Sub-Domain	Data Source	Raw Metric
Clinical	Symptom - Severity	PHQ-9 Questionnaire	Sum of item scores
	Comorbidity	Electronic Health Record	ICD-10 Codes
Biological	Pharmacogene-tics	Genetic Assay	CYP2D6 genotype
	Neuroendo-crinology	Saliva/Blood Sample	Cortisol concentration
Digital Phenotype	Mobility	Smartphone GPS	Latitude/ Longitude
		Smartphone GPS	Latitude/ Longitude
	Social Activity	Smartphone Call Log	Call metadata
		Smartphone SMS Log	SMS metadata
	Circadian Rhythms	Smartphone Screen	Screen on/off timestamps
		Wearable Accelerometer	3-axis acceleration
	Physical Activity	Wearable Accelerometer	Step detection
Domain	Engineered Feature	Sampling Freq.	Theoretical Link to MDD
Clinical	Weekly PHQ-9 Score	Weekly	Core measure of depression severity
	Binary flags for anxiety, SUDs	Static (Baseline)	Comorbidity impacts prognosis
Biological	Poor/Intermediate/ Normal/Ultra-rapid metaboliser status	Static (Baseline)	Influences drug exposure/side effects
	Baseline cortisol level	Static (Baseline)	HPA axis dysregulation
Digital Phenotype	Location Entropy	Daily (from Hz data)	Anhedonia, avolition, behavioural withdrawal
	Time Spent at Home	Daily (from Hz data)	Social withdrawal
	Number of unique contacts	Daily	Social network size/engagement
	Ratio of incoming / outgoing texts	Daily	Social reciprocity
	Inferred sleep duration	Daily	Sleep disturbance, insomnia/hypersomnia
	Circadian Regularity Index	Daily (from Hz data)	Disruption of daily routines
	Daily Step Count	Daily	Psychomotor retardation/agitation

provides a structured data dictionary that defines the variables, sources, and temporal properties comprising the PSV.

Before modelling outcomes, the framework must also formalise the treatment input itself. Complementing the PSV, the Therapeutic Impulse Function (TIF) represents each intervention in a structured, quantitative way that allows the model to simulate how different treatments influence patient trajectories.

2.2. The Therapeutic Impulse Function (TIF): A Formalism for Treatment Inputs

Before modelling outcomes, the framework must also formalise the treatment input itself. Complementing the PSV, the Therapeutic Impulse Function (TIF) captures the structured characteristics of an intervention, enabling quantitative modelling of how different treatments influence recovery trajectories. To test the theory’s proposition that treatment outcomes depend on a specific PSV-TIF interaction, the treatment itself must be represented in a standardised and parameterised format [23]. Simply using a treatment’s brand name as a categorical label discards crucial information about its underlying properties. The TIF represents each intervention as a feature vector describing its pharmacological or procedural properties, transforming treatment selection from a heuristic process into a data-driven problem of system control and optimisation. This reframing allows for a more mechanistic analysis, moving from the question “Which drug is best?” to “For a system in state PSV, what TIF vector will most efficiently guide the system trajectory towards remission?”

• Pharmacotherapy TIF: A vector representation for a medication would include:

  –

  Mechanism of Action: A multi-hot encoded vector representing the primary neurotransmitter systems targeted (e.g., -> for an SNRI).

  –

  Half-life Elimination : A numerical feature for the drug’s elimination half-life (in hours), which governs dosing frequency and time to steady state [30].

  –

  Metabolism: A categorical feature representing the primary CYP450 enzyme responsible for its metabolism (e.g., CYP2D6), allowing for direct modelling of gene-drug interactions.

  –

  Dose: A normalised numerical feature representing the prescribed daily dosage relative to the standard therapeutic range.

• Psychotherapy TIF: A vector representation for a psychotherapy would include:

  –

  Modality: A one-hot encoded vector for the primary therapeutic approach (e.g., Cognitive Behavioural Therapy, Psychodynamic Therapy, Interpersonal Therapy) [25].

  –

  Dose and Schedule: Numerical features for session duration (in minutes) and frequency (sessions per week).

  –

  Process Variables (if available): If data from session ratings is available, a numerical feature for the therapeutic alliance score could be included as a time-varying component of the TIF.

This structured TIF serves as a static or dynamic input to the prognostic model, allowing it to learn not only that treatments differ in outcome, but how and why they do so, based on their intrinsic properties. This representation also facilitates interpretability and integration with decision-support interfaces, where individual treatment parameters can be adjusted and their forecasted effects visualised.

2.3. The Predicted Recovery Trajectory (PRT): Defining the Forecasting Target

The PRT is the primary output and evaluation target of the prognostic framework. Unlike the single-endpoint predictions of static models, the PRT is operationalised as a multi-step sequential forecast representing the expected evolution of symptom severity.

Definition: The PRT is a vector of predicted symptom severity scores (e.g., weekly PHQ-9 scores) over a clinically relevant future horizon, such as 12 weeks. For a model making predictions at week 0, the target output would be [PHQ9 week1, PHQ9 week2, PHQ9 weekX, PHQ9 week12].

Probabilistic Forecasting: A clinically useful forecast must also communicate uncertainty and confidence intervals. Accordingly, the model should output not only a point estimate at each time step but a predictive distribution. This can be achieved through techniques like Monte Carlo dropout or by training the model to predict the parameters (mean and variance) of a Gaussian distribution for each time step. This allows for the visualisation of prediction intervals, giving clinicians a sense of the likely range of future outcomes.

Figure 1 summarises the flow from PSV and TIF to the PRT.

The figure illustrates the flow of information through the proposed methodological system. Multi-modal data streams—including clinical, biological, and digital phenotype domains—are fused into the Patient State Vector (PSV), a dynamic representation of individual state. Treatment characteristics are formalised as the Therapeutic Impulse Function (TIF). These two components jointly serve as inputs to a time-aware Long Short-Term Memory (LSTM) architecture, which forecasts the Predicted Recovery Trajectory (PRT). The PRT output is then embedded within a translational pipeline encompassing interpretability (Explainable AI), ethical governance, and human-centred interface design for clinical decision support. Together, these layers form an integrated framework for developing and deploying prognostic AI in mental health.

3. A Dynamic Forecasting Architecture for the Predicted Recovery Trajectory

With the data structures for the model’s inputs (PSV, TIF) and output (PRT) defined, this section specifies the computational engine of the prognostic framework. Consistent with the theoretical emphasis on temporal dependency and dynamic change, the choice of model architecture is not merely a technical detail; it is a theoretical commitment. A model that cannot process time-series data or capture path-dependent dynamics is fundamentally incompatible with the prognostic theory. This section justifies the selection of a Recurrent Neural Network (RNN) architecture, specifically a time-aware Long Short-Term Memory (LSTM) network with optional attention mechanisms, as the most appropriate tool for this task [48].

3.1. Model Selection Rationale: From State-Space Models to Recurrent Neural Networks

The challenge of modelling psychological and physiological time-series has a long history. A powerful and conceptually elegant approach is the state-space model (SSM) [49,50]. SSMs formalise a system by separating a latent (unobserved) state, which evolves over time according to some underlying dynamic, from the observed measurements, which are manifestations of that latent state. This formalism aligns perfectly with the prognostic theory: the “true” depressive state of a patient is a latent construct, and the various components of the PSV (symptom scores, mobility patterns, etc.) are its noisy, observable indicators. However, classical SSMs often assume linear dynamics and Gaussian noise, which may be overly restrictive for capturing the complex, non-linear processes of psychiatric recovery.

Recurrent Neural Networks (RNNs) can be understood as a powerful, non-linear generalisation of this state-space concept [48,51]. An RNN processes a sequence of inputs one step at a time, maintaining an internal “hidden state” vector that acts as a form of memory. At each time step, the network updates its hidden state based on the current input and its previous hidden state. This hidden state is, in effect, a high-dimensional, learned latent representation of the system’s history, analogous to the state vector in an SSM [39,46,48]. The key advantage is that the complex, non-linear functions governing the state transitions and the mapping from state to observation are learned automatically from the data by the neural network, rather than being pre-specified.

Within the family of RNNs, the Long Short-Term Memory (LSTM) architecture is particularly well-suited for clinical data [18,28,29,48,52,53]. Standard RNNs struggle with the “vanishing gradient” problem, making it difficult for them to learn long-range temporal dependencies. LSTMs overcome this limitation through a series of “gates” (input, forget, and output gates) that explicitly control the flow of information into, out of, and within the cell’s memory. This gating mechanism allows the LSTM to selectively remember important information from the distant past while discarding irrelevant details, a crucial capability for modelling the weeks-long or months-long process of treatment response. The choice of an LSTM, therefore, is not just a technical preference but a computational operationalisation of the theoretical assumption that recovery is path-dependent. Unlike a simple regression model that assumes the future is a direct function of the baseline state, an LSTM’s predictions are conditioned on the entire history of the system as encoded in its hidden state. This structure inherently embodies the theory’s proposition that how a patient arrives at their current state is as important as the state itself [23].

This approach stands in contrast to traditional time-series models like ARIMA, which, while useful for univariate forecasting, are ill-equipped to handle the multi-modal inputs of the PSV and struggle to capture the complex, non-linear dynamics inherent in biological systems. Moreover, the LSTM’s ability to learn temporal dependencies from heterogeneous data makes it particularly suitable for integration into human-centred decision-support systems that require interpretable, time-evolving outputs.

While Transformer-based architectures (e.g., Self-Attention mechanisms) have achieved state-of-the-art results in many time-series domains, we explicitly propose a Time-Aware LSTM for this specific clinical application for two reasons [42]. First, data sparsity: clinical datasets, particularly those with deep phenotyping, rarely reach the sample sizes required to train Transformers without significant overfitting. LSTMs generally demonstrate better inductive bias for smaller, sequence-dependent datasets common in psychiatry [39,54,55,56]. Second, explicit temporal modelling: standard Transformers rely on positional encodings which can struggle with the highly irregular, continuous-time nature of patient visits (e.g., gaps of 2 days vs. 2 months). The Time-Aware LSTM’s gate mechanism (Equation (1)) offers a more direct and biologically plausible method for modelling the decay of information over irregular intervals. Future iterations of this framework may explore “Time-Aware Transformers” as data availability increases [42,46,47].

3.2. Proposed Architecture: A Multi-Input, Time-Aware LSTM for Prognosis

To operationalise the prognostic framework, a multi-input, stacked, time-aware LSTM architecture is proposed. This architecture is specifically designed to address the unique challenges of the PSV and TIF data structures.

Mathematical Formalism of Inputs:

Let the clinical trajectory of a patient be represented by a sequence of observations. We define the model inputs as:

• Dynamic Inputs ($x_t$): A sequence of time-varying vectors $X=(x_1,x_2,\dots ,x_T)$, where $x_t\in {\mathbb{R}}^d$ represents the Dynamic PSV features (e.g., symptom scores, aggregated digital phenotype metrics) at time step t.
• Static Inputs (s): A time-invariant vector $s\in {\mathbb{R}}^k$ representing the Static PSV (genetics, demographics) and the Therapeutic Impulse Function (TIF).
• Time Gaps ($\Delta t$): To account for irregularity, we explicitly calculate the elapsed time between consecutive observations as $\Delta t_t=t_{cur}-t_{prev}$, where t represents the absolute timestamp.

Handling Multi-Modal Inputs:

The model features separate input pathways to accommodate these data types.

• Dynamic Pathway: The sequence $x_t$ is fed into the first LSTM layer to encode temporal dynamics: $h_t^{\left(1\right)}={\mathrm{LSTM}}_1\left(x_t,h_{t-1}^{\left(1\right)}\right)$.
• Static Pathway: The static vector s is concatenated with the hidden state output of the first layer (⊕ denotes concatenation) to condition the high-level feature learning: $u_t=\left[h_t^{\left(1\right)}\oplus s\right]$. This fused vector $u_t$ serves as the input to the second LSTM layer: $h_t^{\left(2\right)}={\mathrm{LSTM}}_2(u_t,h_{t-1}^{\left(2\right)})$.

Time-Awareness:

A critical limitation of standard LSTMs is that they treat observations as if they occur at regular time intervals, ignoring the actual time gaps $\Delta t_t$. To overcome this, the proposed architecture incorporates explicit time-decay gates into the LSTM cell, inspired by the KIT-LSTM model [45].

Standard LSTMs update the cell memory $c_t$ using an input gate $i_t$, forget gate $f_t$, and candidate memory ${\tilde{C}}_t$. We modify this by introducing a time-decay gate $g_t$ that is a monotonic decreasing function of the time gap:

$$g_t=\sigma \left({\displaystyle \frac{1}{log(e+\Delta t_t)}}\right)$$

(1)

This gate acts on the previous cell memory $c_{t-1}$ to discount information based on the elapsed time, creating an adjusted memory state $c_{t-1}^{*}$ before the standard LSTM update occurs:

$$c_{t-1}^{*}=c_{t-1}\odot g_t$$

(2)

where ⊙ denotes element-wise multiplication. This mechanism allows the model to “forget” more of its memory as the time gap $\Delta t_t$ increases, ensuring that a symptom score from yesterday is weighted more heavily than one from two weeks ago. This step performs a temporal correction on the historical memory, allowing recent clinical states to retain a stronger influence while attenuating information originating from older or irregularly spaced observations.

Using this time-adjusted memory, the final LSTM cell-state update follows the standard formulation, with the forget gate $f_t$ input gate $i_t$, and candidate memory ${\tilde{C}}_t$:

$$c_t=f_t\odot c_{t-1}^{*}+i_t\odot {\tilde{c}}_t.$$

(3)

Together, these two equations (Equations (2) and (3)) constitute the complete Time-Aware LSTM memory mechanism. The first equation embeds sensitivity to temporal irregularity, while the second integrates the decayed historical information with new clinical input. This sequential formulation preserves the interpretability and gating structure of the classical LSTM while enabling principled modelling of PSV trajectories with inconsistent sampling intervals.

Attention Mechanism (Extension):

For enhanced performance and interpretability, the architecture can be extended with a temporal attention mechanism. The model computes a context vector z as a weighted sum of past hidden states $h_i$:

$$z=\sum _{i=1}^T{\alpha }_i{h}_i$$

(4)

where ${\alpha }_i$ are attention weights derived from a learned compatibility function. This allows the model to dynamically focus on specific historical periods (e.g., weeks 2–4) when forecasting the PRT, directly addressing the theory’s proposition about the importance of early response dynamics [23].

Interpretability and Clinical Integration:

To align with the human-centred orientation of this work, the modular structure allows for transparent analysis. The attention weights ${\alpha }_i$ provide a direct measure of temporal saliency, indicating which time points most influenced a specific prognosis. Similarly, gradient-based attribution methods can be applied to the static input vector s to quantify the contribution of specific TIF parameters (e.g., drug dosage) or biological markers to the predicted trajectory.

3.3. Training, Validation, and Evaluation Protocol

A rigorous and transparent protocol for model development and evaluation is essential for ensuring the reliability and replicability of the results. The focus of this protocol must shift from optimising for the best endpoint classification accuracy to achieving the most clinically informative and accurate trajectory forecast.

• Data Preprocessing: All continuous input features from the PSV and TIF will be standardised (e.g., z-score normalisation) to ensure they are on a comparable scale, a standard practice for training neural networks.
• Training: The model will be trained end-to-end using backpropagation through time (BPTT). The Adam optimiser, an adaptive learning rate optimisation algorithm, will be used to minimise the loss function. The primary loss function will be the Root Mean Squared Error (RMSE) calculated across all time points of the predicted and true trajectories.
• Validation: To prevent information leakage and ensure the model generalises to unseen patients, a patient-level, nested cross-validation scheme will be employed. The outer loop splits patients into training and test sets. The inner loop performs hyperparameter tuning (e.g., number of LSTM units, learning rate, dropout rate) on a validation set carved out from the training set. This approach ensures patient-level independence and provides a robust estimate of model generalisability.
• Evaluation Metrics: The evaluation must reflect the prognostic goal of forecasting the entire trajectory. While traditional classification metrics can be reported for benchmarking, the primary metrics should be:

  –

  Trajectory-wise Root Mean Squared Error (RMSE) or Mean Absolute Error (MAE): The average error calculated across all forecasted time points in the PRT. This is the primary measure of the model’s ability to accurately predict the entire path of recovery.

  –

  Endpoint MAE: The absolute error at the final time point of the forecast horizon (e.g., week 12). This allows for direct comparison with static models that only predict this single point.

  –

  Dynamic Time Warping (DTW): A metric that measures the similarity between two temporal sequences that may vary in speed. It is particularly useful for assessing whether the model has captured the correct shape of the recovery trajectory (e.g., rapid initial response followed by a plateau), even if it is slightly off in its timing.

To demonstrate the value of this dynamic, multi-modal approach, the proposed model must be benchmarked against a series of simpler models.

Table 2. Model Evaluation Benchmarks.

Model	Input Data	Evaluation Metrics
Multiple Linear Regression (Baseline)	Baseline PSV only (static features)	Endpoint MAE, AUC at 12 weeks
ARIMA	Univariate time-series of clinical scores only	Trajectory RMSE, Endpoint MAE
Standard LSTM	Full time-series of PSV + TIF (no time-awareness)	Trajectory RMSE, Endpoint MAE, DTW
Time-Aware LSTM (Proposed)	Full time-series of PSV + TIF (with time-awareness)	Trajectory RMSE, Endpoint MAE, DTW

outlines this evaluation framework.

This structured comparison will allow for a clear quantification of the performance gains attributable to (a) using a dynamic model over a static one, (b) incorporating multi-modal data over univariate data, and (c) explicitly modelling irregular time intervals. This presents a rigorous, evidence-based argument in support of the proposed methodology.

In addition to quantitative performance, qualitative inspection of trajectory shapes and attention maps will provide evidence for clinical plausibility and interpretability—key criteria for human-centred AI in psychiatry.

Proposed Synthetic Validation Protocol

To validate the architectural advantages of the Time-Aware LSTM prior to clinical deployment, we propose a synthetic data experiment designed to isolate the effects of temporal irregularity.

• Data Generation: We generate synthetic patient trajectories $y\left(t\right)$ using a damped harmonic oscillator equation with injected noise, representing the cyclical but decaying nature of mood episodes.
• Irregular Sampling Simulation: We sample this ground-truth trajectory at random intervals $\Delta t\sim \mathrm{Exponential}\left(\lambda \right)$, creating a “sparse” observation set typical of clinical data.
• Hypothesis Testing: We test the hypothesis that standard LSTMs will fail to capture the recovery rate because they treat the random $\Delta t$ as constant steps. The Time-Aware LSTM, using the gate $g_t=\sigma \left({\displaystyle \frac{1}{log\left(e+\Delta t\right)}}\right)$, is expected to recover the underlying decay parameter despite the irregular sampling.

This protocol allows for the verification of the “Time-Aware” mechanism (Equations (1)–(3)) independent of clinical noise, serving as a necessary gate before training on expensive real-world data.

4. Bridging the Translational Gap: A Blueprint for Responsible Implementation

A prognostic model, no matter how accurate, is clinically inert if it is untrusted, unusable, or unethical. The history of AI in medicine is littered with technically successful models that failed to translate into practice because these crucial human-centred factors were ignored [23]. Therefore, a complete methodological framework must not end with model validation; it must extend to include a proactive blueprint for responsible implementation. The deployment of a prognostic model creates a new form of socio-technical-ethical debt. The act of collecting the highly sensitive PSV incurs a debt to patient privacy; the act of generating a high-stakes forecast incurs a debt of accountability. These interdependent debts highlight that model performance and ethical integrity are inseparable dimensions of translational success. The ultimate success metric for this framework is not simply predictive accuracy, but clinical adoption and stakeholder trust, which are fundamentally outcomes of good design, ethical integrity, and clear communication, as highlighted by the phases in

Table 3. Implementation Roadmap: Phases, Risks, and Mitigations.

Implementation Phase	Key Deliverable	Primary Risk	Methodological Mitigation
Phase 1: Data Architecture	Constructed PSV & TIF vectors.	Sparsity & Noise: Sensors fail; patients skip surveys.	Impute missingness via time-decay gates; use “masking” layers in LSTM training.
Phase 2: Model Training	Trained Time-Aware LSTM.	Overfitting: Model memorises specific patient histories.	Nested Cross-Validation (patient-level split); Regularisation (Dropout).
Phase 3: Explainability (XAI)	SHAP plots & Attention Maps.	Misinterpretation: Clinicians over-rely on false signals.	Counterfactual testing (“What if?”); Uncertainty quantification (Confidence Intervals).
Phase 4: Clinical Integration	Decision Support Interface.	Alert Fatigue: Too many false alarms.	Set high specificity thresholds; co-design interface with clinicians (HCD).
Phase 5: Governance	Bias Audit Report.	Algorithmic Bias: Model fails for minority groups.	Mandatory performance disaggregation by demographic group; Federated Learning for privacy.

.

4.1. Ensuring Clinical Trust: From Black Box to Interpretable Prognosis (XAI)

The “black box” nature of deep learning models is one of the most significant barriers to their clinical adoption [23,57,58]. For a clinician to act on a model’s prediction—especially one that may contradict their own judgement—they must have a plausible and understandable rationale for that prediction [23]. This requires integrating Explainable AI (XAI) techniques into the core methodology. Interpretability is not a single feature but a multi-level requirement that must be tailored to the needs of different stakeholders: developers need technical diagnostics, clinicians need actionable clinical narratives, and patients need empowering insights. In this context, interpretability serves not only epistemic transparency but also epistemic responsibility, the obligation to make predictions that can be justified and communicated within clinical reasoning frameworks.

• Feature Importance (Global and Local):

  –

  Technique: Methods like SHAP (Shapley Additive Explanations) can be used to quantify the contribution of each input feature to a specific prediction.

  –

  Clinical Application: After the model forecasts a PRT of non-response for a patient, SHAP can reveal the primary drivers of this prediction. For example, it might highlight that persistent sleep disruption (from the digital phenotype) and high baseline inflammatory markers are the top contributing factors. This transforms an opaque prediction (“non-responder”) into an interpretable clinical narrative (“The model predicts non-response, likely due to unresolved sleep and inflammation issues”), providing the clinician with a testable hypothesis to guide their next steps.

• Temporal Saliency:

  –

  Technique: If an attention mechanism is used, the attention weights themselves can be visualised to show which past time points the model focused on when making its forecast. Alternatively, gradient-based saliency methods can be used.

  –

  Clinical Application: This can identify critical periods in a patient’s illness course. For instance, the model might learn that a small dip in mood and social activity during week 3, even if it recovers, is a strong predictor of eventual relapse. This provides clinicians with insight into the importance of early, subtle fluctuations that might otherwise be dismissed. Such outputs can be visualised directly in the interface, allowing clinicians to inspect which temporal dynamics contributed most to a forecast—enhancing both trust and educational value.

• Counterfactual Explanations:

  –

  Technique: These methods generate “what-if” scenarios by minimally perturbing the input to change the model’s output.

  –

  Clinical Application: This provides the most actionable form of explanation. The system could answer questions like, “How would the predicted trajectory change if this patient’s daily step count increased by 2000?” or “What is the minimum improvement in sleep regularity needed to shift the forecast from non-response to response?” This bridges model output with behavioural and clinical levers of change, aligning algorithmic insight with therapeutic reasoning. This directly connects the model’s prediction to modifiable behaviours and helps in collaborative treatment planning between the clinician and patient.

4.2. An Ethical Framework for Prognostic AI

The collection and use of data for the PSV, particularly the continuous and intimate data from digital phenotyping, raises profound ethical challenges that must be addressed proactively [23,59]. A robust ethical framework is not an optional add-on but a core component of the methodology. Ethics here is treated as a continuous design principle, not as a compliance checklist.

• Data Privacy and Informed Consent:

  –

  Challenge: The passive, continuous nature of digital phenotyping is incompatible with traditional, one-time consent models [40,41]. Patients must have an ongoing understanding and control over what data is being collected and how it is being used [60,61,62].

  –

  Methodological Solution: Implementation of a dynamic consent interface within the data collection application. This would allow patients to view the data being collected, understand its purpose in simple terms, and have granular control to pause or withdraw specific data streams at any time. Furthermore, privacy-preserving machine learning techniques, such as federated learning, should be explored. In this approach, the model is trained on the user’s device, and only the updated model weights, not the raw personal data, are sent to a central server, significantly enhancing privacy [63]. Complementary approaches, such as differential privacy and secure multiparty computation, could further minimise re-identification risks in aggregated datasets.

• Algorithmic Bias and Fairness:

  –

  Challenge: AI models trained on historical data can learn and amplify existing societal biases and health disparities [23,64,65,66,67,68]. A model trained predominantly on one demographic group may perform poorly and unfairly on underrepresented populations, exacerbating inequities in care [69,70,71,72].

  –

  Methodological Solution:A mandatory algorithmic bias audit must be part of the model validation protocol. This involves disaggregating model performance metrics (e.g., Trajectory RMSE) across key demographic subgroups (e.g., race, gender, age, socioeconomic status) [73,74]. If significant performance disparities are found, mitigation strategies must be employed, such as re-weighting the training data, applying fairness constraints during training, or collecting more data from the underperforming subgroup [73,74]. Results from these audits should be reported transparently in Supplementary Materials, establishing accountability and reproducibility.

• Accountability and Responsibility:

  –

  Challenge: If the model’s forecast contributes to an adverse patient outcome, who is responsible? The developer, the hospital, or the clinician who used the tool? [75].

  –

  Methodological Solution: The framework must establish clear lines of accountability. The model must be legally and ethically framed as a Clinical Decision Support (CDS) tool, not a medical device that makes autonomous decisions. The final clinical judgement and responsibility must always reside with the human clinician. The system’s documentation and user interface must explicitly state its probabilistic nature, its limitations, and its role as an assistive tool to augment, not replace, professional expertise. Explicit audit logs should track how model outputs are used in clinical decisions, supporting traceability and learning from errors [75,76,77].

• Regulatory Compliance:

  –

  Challenge: The legal and regulatory landscape for AI in healthcare is rapidly evolving, with new guidelines from bodies like the WHO and FDA, and new laws at the state and national levels [73,74,78].

  –

  Methodological Solution: The development process must incorporate a regulatory monitoring step, ensuring that data handling practices comply with existing regulations like HIPAA and GDPR, and that the tool’s classification (e.g., as a wellness app vs. a medical device) is appropriate and defended [59,73]. A risk-based framework should be adopted, where the level of regulatory scrutiny is proportional to the potential impact of the tool’s output on patient care [64,65,66,67,68].

This anticipatory governance approach ensures the model evolves in compliance with emerging ethical and legal norms rather than reacting post hoc.

Additional considerations for sensing-based prognostic systems: Within the context of mobile and wearable sensing, ethical considerations must also address the continuous and often unobtrusive collection of behavioural signals, including location traces and device-level identifiers. Key challenges involve managing real-time consent, deciding between local versus cloud-based processing, ensuring secure on-device storage, and applying differential privacy mechanisms to long-term behavioural streams. These issues differ from traditional clinical ethics and are essential for the responsible deployment of any prognostic framework that relies on sensor-derived data.

4.3. Human-Centred Design for Clinical Integration (HCI)

Even a perfectly accurate, interpretable, and ethical model will fail if it is not designed to fit seamlessly and usefully into the complex realities of clinical workflow [79,80]. The interface through which clinicians and patients interact with the prognostic forecasts is a critical component of the overall system. A human-centred design (HCD) process is therefore a foundational part of the methodology. By embedding usability testing and participatory design cycles throughout model development, the system’s interface becomes a co-evolving element of the methodology itself.

• User-Centred Design Process:

  –

  Challenge: CDS tools are often designed by engineers with little understanding of clinical realities, leading to poor workflow integration and low adoption rates [75].

  –

  Methodological Solution: The design process must be iterative and participatory, involving clinicians, patients, and administrators from the earliest stages. Techniques like semi-structured interviews, workflow analysis (journey mapping), and the development of user personas can be used to deeply understand the needs, goals, and pain points of the end-users [75]. This ensures the final tool solves a real problem in a way that minimises, rather than increases, the clinician’s cognitive load. Iterative prototyping and usability metrics (e.g., task completion time, NASA-TLX workload) should be systematically recorded to ensure continuous alignment with user needs.

• Designing for Generalisability and Fairness:

  –

  Cross-Population Generalisation and Fairness: Ensuring fairness in prognostic AI extends beyond mitigating algorithmic bias during training. It also requires systematic evaluation of model generalisation across institutions, devices, and demographic groups. Performance audits should therefore be conducted across independent clinical sites and heterogeneous populations to identify and correct any degradation in accuracy or interpretability. Such cross-site validation not only strengthens the model’s robustness but also ensures equitable clinical benefit, preventing the amplification of existing disparities in access or outcomes.

• Interface Design for Probabilistic Information:

  –

  Challenge: Presenting a single, deterministic trajectory forecast can create a false sense of certainty and undermine clinical judgement.

  –

  Methodological Solution: The interface must be designed to communicate uncertainty effectively. The primary visualisation should not be a single line, but a “cone of probability,” showing the mean predicted trajectory surrounded by shaded confidence intervals (e.g., 50% and 95% prediction intervals). The interface should also allow for interactive exploration, enabling the user to view the counterfactual explanations (“What if we change the TIF to this other medication?”) and see how the probabilistic forecast shifts in response. This interactivity is not cosmetic; it transforms the forecast into a shared reasoning space where clinicians can simulate and discuss treatment alternatives transparently. Simplicity, clarity, and effective use of colour and layout are paramount to ensure the information is glanceable and interpretable in a busy clinical setting [79,80].

• Supporting the Therapeutic Alliance:

  –

  Challenge: There is a significant risk that an AI tool could be perceived as an impersonal, algorithmic authority, thereby disempowering the patient and eroding the human connection at the heart of therapy [59,75].

  –

  Methodological Solution: The CDS tool should be designed explicitly to facilitate a collaborative conversation. It can be used as a shared visual aid during a consultation. A clinician could show a patient their own digital phenotype data (e.g., “You can see here how your sleep regularity has improved over the last two weeks”) and then connect it to the forecasted PRT (“The model suggests this improvement is a very positive sign for your long-term recovery”). This reframes the tool from a top-down predictor to a bottom-up facilitator of shared understanding and self-efficacy, leveraging the patient’s own data to empower them in their recovery process. It supports the therapeutic relationship by providing a common, data-driven ground for discussion and goal-setting.

• Illustrative Use Case: To illustrate the envisioned clinical application, consider a follow-up consultation in which a clinician uses the prognostic decision support interface with a patient. The model forecasts a flattening in the predicted recovery trajectory over the next four weeks and highlights disrupted sleep regularity and reduced mobility as primary contributing factors. Guided by this insight, the clinician and patient review recent behavioural data, discuss possible stressors, and agree on specific adjustments to improve daily routines and sleep hygiene. This interaction exemplifies how the proposed system can function as an interpretive, collaborative aid—supporting clinical reasoning and shared understanding rather than replacing human judgement.

Ultimately, the tool’s success should be evaluated not only by predictive metrics but by its measurable contribution to communication quality, patient engagement, and therapeutic alliance strength.

5. Discussion

5.1. Summary of the Methodological Contribution

This methodological contribution has presented a comprehensive, end-to-end methodological framework for operationalising a prognostic theory of mental health treatment response. Its purpose is to provide the necessary conceptual and technical blueprint to move the field from the prevailing paradigm of static, endpoint-based prediction toward a more clinically relevant paradigm of dynamic, trajectory-based prognosis. By uniting theoretical constructs, computational architectures, and human-centred design principles, this framework articulates a coherent research agenda for prognostic psychiatry. The key contributions of this framework are threefold.

First, it provides a granular and practical guide for operationalising the core theoretical constructs of the prognostic theory. It details a multi-modal, high-frequency data architecture for the Patient State Vector (PSV), a formal vector representation for the Therapeutic Impulse Function (TIF), and defines the Predicted Recovery Trajectory (PRT) as a suitable, multi-step forecasting target. This moves the theory from an abstract concept to an empirically testable reality.

Second, it proposes and justifies a theoretically grounded computational architecture for forecasting. The selection of a time-aware Long Short-Term Memory (LSTM) network is not arbitrary but is rooted in the dynamic, path-dependent nature of the recovery process as conceptualised by the theory. The architecture is specifically designed to handle the multi-modal, asynchronous, and irregularly-sampled data that characterises real-world clinical research. It thus translates theoretical principles of temporal dependency and individual variability into a computational form that can be empirically interrogated.

Third, and most critically, it argues that a viable methodology cannot be purely technical. It integrates a holistic translational pipeline as a core, non-negotiable component. This includes a protocol for ensuring model interpretability via XAI, a governance framework for navigating the profound ethical challenges of prognostic AI, and a set of human-centred design principles for creating a clinical decision support tool that is usable, trustworthy, and supportive of the therapeutic alliance. In positioning interpretability, ethics, and usability as methodological pillars rather than afterthoughts, the framework embodies the interdisciplinary ethos required for responsible AI in mental health and proactively addresses the primary reasons why promising computational models so often fail to impact patient care.

5.2. Limitations and Future Directions

Despite its comprehensive nature, this proposed framework has significant limitations that point toward important avenues for future research. The most formidable challenge is the immense data requirement. Assembling the full PSV, with its integration of clinical, biological, and high-frequency digital phenotype data, is a logistically complex and resource-intensive undertaking. Few, if any, existing datasets contain all of these components collected longitudinally in a large cohort. This limitation underscores the call made in the original theoretical paper for the establishment of large-scale, openly shared, multi-modal longitudinal cohorts, which are an essential prerequisite for training and validating robust prognostic models [18,23]. Future collaborations between clinical consortia and digital health platforms may provide the infrastructure necessary to meet this challenge.

On the computational front, while LSTMs are well-suited for capturing temporal dependencies, they can be computationally expensive to train and may struggle to capture very long-range dependencies spanning many months. Future work should explore more recent and potentially more powerful sequence-modelling architectures, such as Transformers, which have shown state-of-the-art performance in other time-series domains [43,44,81]. The attention mechanism at the core of the Transformer architecture could be particularly adept at identifying complex, non-linear interactions across a patient’s entire history. Moreover, hybrid architectures combining LSTMs and Transformers may balance interpretability with scalability, offering a pragmatic bridge toward clinical deployment.

Furthermore, the framework presented here focuses on a single-disease model (MDD). The next evolution of this work will involve extending the methodology to transdiagnostic applications, modelling the dynamic interplay between different symptom domains (e.g., depression, anxiety, sleep) and forecasting trajectories across multiple dimensions of mental health simultaneously. Such extensions could pave the way toward a unified computational taxonomy of psychiatric prognosis, aligning with emerging dimensional frameworks like the Research Domain Criteria (RDoC).

A critical next step for this line of research is the empirical validation of the proposed framework in real-world clinical contexts. Future studies should adopt a staged evaluation strategy, beginning with retrospective validation using existing longitudinal datasets, followed by prospective observational studies, and ultimately clinician-in-the-loop trials. Such a phased approach would enable the assessment of predictive performance, interpretability, and usability in parallel, ensuring that methodological advances translate into genuine clinical value. Moreover, deployment studies within clinical decision support environments will be essential to evaluate the framework’s capacity to augment clinician reasoning, adapt to diverse workflows, and sustain ethical and transparent operation under real-world constraints.

5.3. Conclusion: A Roadmap for Prognostic Psychiatry

This methodological contribution has laid out a methodological roadmap. It is a bridge designed to connect the abstract potential of a new theory to the reality of empirical validation and, ultimately, clinical application. The trial-and-error nature of psychiatric care is not an intractable problem but a scientific challenge that demands new tools and new ways of thinking. The shift from static prediction to dynamic prognosis is a necessary evolution for the field. The framework proposed here represents a step toward this transformation by defining how data, computation, and human factors can converge into a coherent system of prognostic reasoning.

By adopting a comprehensive, rigorous, and ethically-aware methodology—one that integrates sophisticated data architectures, theoretically-grounded models, and a robust translational pipeline- the research community can begin to build and validate the tools needed to realise the vision of a truly personalised and prognostic psychiatry. The impact of this framework should ultimately be measured not by technical novelty alone, but by its capacity to inform real-world clinical decision-making, enhance patient engagement, and improve treatment outcomes. This framework is not an endpoint, but a starting point for the empirical work ahead, a call to action to build the data infrastructure and computational systems that will finally allow the field to ensure that every patient has the best chance of receiving the right treatment, at the right time, on the right trajectory toward recovery.