Advancing Precision Neurology and Wearable Electrophysiology: A Review on the Pivotal Role of Medical Physicists in Signal Processing, AI, and Prognostic Modeling

Abstract

Medical physicists are transforming physiological measurements and electrophysiological applications by addressing challenges like motion artifacts and regulatory compliance through advanced signal processing, artificial intelligence (AI), and statistical rigor. Their innovations in wearable electrophysiology achieve 8–12 dB signal-to-noise ratio (SNR) improvements in EEG, 60% motion artifact reduction, and 94.2% accurate AI-driven arrhythmia detection at 12 μW power. In precision neurology, machine learning (ML) with evoked potentials (EPs) predicts spinal cord injury (SCI) recovery and multiple sclerosis (MS) progression with 79.2% accuracy based on retrospective data from 560 SCI/MS patients. By integrating multimodal data (EPs, MRI), developing quantum sensors, and employing federated learning, these can enhance diagnostic precision and prognostic accuracy. Clinical applications span epilepsy, stroke, cardiac monitoring, and chronic pain management, reducing diagnostic errors by 28% and optimizing treatments like deep brain stimulation (DBS). In this paper, we review the current state of wearable devices and provide some insight into possible future directions. Embedding medical physicists into standardization efforts is critical to overcoming barriers like quantum sensor power consumption, advancing personalized, evidence-based healthcare.

1. Introduction

Physiological measurements, including wearable electrophysiology and evoked potentials (EPs), are pivotal for diagnostics and personalized medicine, enabling continuous monitoring and prognostic assessment. Wearable electrocardiography (ECG) and electroencephalography (EEG) face motion artifacts (50–500 μV noise), while EPs for SCI and MS require precise prognostic tools [1,2]. Emerging technologies like wearable photonic smart wristbands further expand capabilities, using photonic sensors to monitor an individual’s respiratory rate (RR), heart rate (HR), blood pressure (BP), and biometric identification with high precision and low power consumption [3]. Medical physicists address these challenges through expertise in signal processing, AI, uncertainty quantification, and statistical validation, delivering reliable data on arrhythmias, epilepsy, and neurological disorders [4]. Their work achieves 8–12 dB SNR improvements in EEG, 94.2% accurate AI-driven diagnostics, and 79.2% accurate EP-based ML models for SCI/MS prognosis [2]. This article explores their contributions, advocating for their integration into device development and standardization to advance precision neurology and wearable medical technology. The integrated role of medical physicists in precision neurology and wearable electrophysiology is summarized in the study workflow, from signal acquisition to clinical applications (Figure 1).

2. Pathophysiological Framework: Linking EPs to Neurological Disorders

SCI and MS impair neural conduction through demyelination, inflammation, and axonal damage, making EPs critical biomarkers [5,6]. In SCI, trauma induces oligodendrocyte apoptosis, reducing myelin and prolonging EP latencies (e.g., N20, P40), while inflammation (TNF-α, IL-6) reduces amplitudes [7]. Axonal transection results in absent amplitudes, signaling irreversible loss [5]. In MS, autoimmune myelin attacks cause multifocal demyelination, detected by prolonged SSEP latencies, with chronic inflammation reducing MEP amplitude [8]. Impaired remyelination leads to persistent EP abnormalities [6,9]. Medical physicists leverage EPs to capture these processes, enabling ML models to predict SCI recovery (ASIA scale) and MS progression (EDSS) with 79.2% accuracy [2].

3. The Expertise of Medical Physicists and Biophysicists

Medical physicists and biophysicists integrate physics, mathematics, biology, statistics, and computational intelligence to develop validated physiological measurement technologies [10]. They transform noisy biosignals into clinical insights, spanning wearable ECG-, EEG-, and EP-based diagnostics emphasizing gaps in real-time prognostic tools.

In wearable electrophysiology, they achieve 8–12 dB SNR improvements in EEG using wavelet-based denoising, outperforming conventional filters by 2–4 times [11]. Ag/AgCl nanostructured dry electrodes reduce impedance from 50–100 kΩ to <5 kΩ, enabling reliable ambulatory monitoring [12]. Biophysicists develop 1D Convolutional Neural Networks (CNNs), achieving 94.2% accuracy in arrhythmia detection at 12 μW, supporting week-long monitoring on a 100mAh battery [13]. In precision neurology, they design EP-based ML models, integrating SSEP latency and MEP time series to predict SCI/MS outcomes with 79.2% accuracy [2]. Their multidisciplinary expertise ensures precision and clinical relevance.

3.1. Precision in Measurement Techniques

Precision is critical in electrophysiological applications, where subtle signal variations indicate conditions like seizures, arrhythmias, or neurological deficits [10]. Medical physicists employ wavelet packet decomposition (WPD) with six-level Daubechies-4 transforms and adaptive thresholding (Tj = σj√(2lnNj)), reducing motion artifacts by 60% compared to finite impulse response (FIR) filters (p < 0.01) [11]. This enables ambulatory EEG and ECG monitoring, detecting epileptiform spikes or arrhythmic patterns with high fidelity.

Nanostructured Ag/AgCl dry electrodes lower noise floors from 5–10 μV to 0.8–1.2 μV, supporting 72 h epilepsy monitoring with <5% data loss versus 30–40% in conventional systems [12]. Calibration to IEC 60601-2-47 ensures RR-interval uncertainties of ±2.8 ms, outperforming commercial devices (±5 ms) [14]. For EPs, precise SSEP latency measurements (e.g., N20, P40) correlate with demyelination, enhancing prognostic accuracy in SCI/MS [2]. We posit that medical physicists’ expertise in signal processing and AI can overcome SNR and artifact barriers in wearable electrophysiology.

3.2. Advantages and Disadvantages of Wavelet Packet Decomposition

Wavelet packet decomposition (WPD) offers significant advantages for EEG/ECG signal processing compared to finite impulse response (FIR) filters and independent component analysis (ICA). Advantages include superior time–frequency localization, enabling precise isolation of motion artifacts (0.5–4 Hz) from neural signals (8–30 Hz), achieving 8–12 dB SNR improvements versus 3–5 dB for FIR and 3.2 dB for ICA in EEG denoising [11,15]. WPD’s adaptive nature outperforms ICA’s assumption of source independence, reducing artifacts by 60% in dynamic ambulatory environments, critical for wearable applications [16]. Unlike FIR’s linear filtering, which smears non-stationary transients, WPD preserves epileptiform spikes and QRS complexes, enhancing diagnostic fidel [17]. Disadvantages include higher computational complexity (O(n log n) vs. FIR’s O(n)), requiring optimized hardware for real-time use, and parameter sensitivity (e.g., wavelet type, decomposition levels), where improper tuning risks over-decomposition and signal distortion [18].

Our contribution advances WPD by optimizing the Daubechies-4 wavelet with a six-level decomposition tailored for EEG, reducing processing time by 30% (from 1.2 s to 0.84 s per 10 s epoch) while maintaining 8–12 dB SNR gains. This optimization, achieved through adaptive thresholding ((Tj =\σj\sqrt{2\ln Nj})), enhances real-time performance on resource-constrained wearables, bridging gaps in ambulatory epilepsy and cardiac monitoring. Compared to ICA, which struggles with correlated artifacts (e.g., eye blinks, motion), our WPD implementation isolates low-frequency noise with 15% higher accuracy in spike detection [2]. In ECG, WPD’s multi-resolution analysis captures arrhythmic patterns with 94.2% accuracy, outperforming FIR’s 88% by preserving transient features [19]. Future work will integrate WPD with edge-AI to further reduce latency to <0.5 s, enhancing clinical translation. Signal processing techniques, such as wavelet packet decomposition (WPD), achieve significant SNR improvements (Figure 2).

3.3. Wavelet Packet Decomposition Implementation

WPD decomposes EEG signals into sub-bands, isolating motion artifacts (0.5–4 Hz) from neural signals (8–30 Hz). The algorithm uses a Daubechies-4 wavelet, applying six-level decomposition and adaptive thresholding to reconstruct artifact-free signals. The process achieves 8–12 dB SNR improvements, critical for epilepsy monitoring (Algorithm 1).

The WPD coefficient calculation is as follows:

Algorithm 1: Wavelet Packet Decomposition

[ c_{j,k} = \int s(t) \psi_{j,k}(t), dt ] where (\psi_{j,k}(t)) is the wavelet function scaled and shifted at level (j) and position (k).

The adaptive threshold is as follows:

[ T_j = \sigma_j \sqrt{2 \ln N_j} ] where (\sigma_j) is the sub-band variance and (N_j) is the length. Soft thresholding is applied as: [ W_{j,k} = \sign(W_{j,k}) \cdot \max(|W_{j,k}| − T_j, 0) ] Algorithm: Wavelet Packet Decomposition for EEG Denoising Input: Raw EEG signal (x(t)), sampling rate (f_s) Output: Denoised EEG signal (x_d(t)) Initialize Daubechies-4 wavelet (db4)

Perform 6-level WPD: decompose (x(t)) into sub-bands (W_{j,k}) For each sub-band (j,k): a. Compute variance (\sigma_j = \sqrt{\text{mean}(W_{j,k}^2)}) b. Set threshold (T_j = \sigma_j \sqrt{2 \ln N_j}) where (N_j) is sub-band length c. Apply soft thresholding: (W_{j,k} = \sign(W_{j,k}) \cdot \max(|W_{j,k}| − T_j, 0)) Reconstruct signal (x_d(t)) using inverse WPD Return (x_d(t))

3.4. Adaptive Noise Cancellation for ECG

Adaptive noise cancellation (ANC) removes motion artifacts in ECG using a reference signal (e.g., accelerometer data). The adaptive filter minimizes the error (e(n) = s(n) − y(n)), where (s(n)) is the noisy ECG and (y(n)) is the filtered noise estimate, achieving 70% artifact reduction (Algorithm 2), [20].

The error minimization is as follows:

Algorithm 2: Adaptive Noise Cancellation

[ e(n) = s(n) − y(n) = s(n) − \sum_{i = 0}^{M − 1} w_i(n) r(n − i) ] Weight update using LMS algorithm: [ w_i(n + 1) = w_i(n) + \mu e(n) r(n − i) ] Algorithm: Adaptive Noise Cancellation for ECG Input: Noisy ECG (s(n)), reference noise (r(n)) Output: Clean ECG (e(n))

Initialize filter weights (w(0) = 0) For each sample (n): a. Compute filter output: (y(n) = w(n) ∗ r(n)) b. Compute error: (e(n) = s(n) − y(n)) c. Update weights: (w(n + 1) = w(n) + \mu ∗ e(n) ∗ r(n)) Return (e(n))

3.5. Uncertainty Calculations and Statistical Rigor

Machine learning (ML) methods are essential for processing complex physiological data, enabling predictive modeling in neurology. Random Forest (RF) is an ensemble method that builds multiple decision trees, averaging predictions to reduce overfitting; it is commonly used for EP-based prognosis, with high accuracy (83–84.7%) in SCI/MS [2]. Support Vector Machines (SVM) classify data by finding hyperplanes that maximize margins; they are effective for high-dimensional EP time series [21]. Deep Learning, including Convolutional Neural Networks (CNNs), extracts features from signals like ECG, achieving 94.2% arrhythmia detection [22]. Neural Networks (NN) learn non-linear relationships, predicting disability progression with 98% accuracy in MS [23]. These methods, combined with cross-validation, ensure robust performance, but require feature selection to avoid bias.

3.5.1. Reasons for 94.2% Accuracy in AI-Driven Arrhythmia Detection

The 94.2% accuracy in AI-driven arrhythmia detection is achieved through several synergistic factors. The 1D Convolutional Neural Network (CNN) architecture, optimized with 32 and 64 filters (kernel sizes 5 and 3), efficiently extracts temporal features from ECG signals, capturing subtle patterns like R-wave variations and QRS complex morphology in the MIT-BIH Arrhythmia Database (n = 47,000 samples, five classes: normal, ventricular ectopic, supraventricular ectopic, fusion beats, unknown) [22]. Model quantization to 68KB using 8-bit integer operations reduces computational overhead while preserving performance, enabling real-time processing at 12 μW power on a 100mAh battery for week-long monitoring. Robust training with 5-fold cross-validation minimizes overfitting, achieving 94.2% accuracy (95% CI: 93.8–94.6%) on test sets, with a low false positive rate (5.1% vs. 8–10% in threshold-based methods). High-quality ECG preprocessing, using adaptive noise cancellation (ANC) for 70% artifact reduction, ensures clean input signals [24]. Compared to traditional methods (e.g., threshold-based detection, 80–85% accuracy; Pan–Tompkins algorithm, 88% accuracy), the CNN’s hierarchical feature learning and ability to model non-linear patterns enhance reliability for wearable applications [25]. Limitations include potential overfitting to specific arrhythmias (e.g., ventricular ectopic beats) and reduced accuracy in noisy ambulatory settings (90–92% in high-motion scenarios), mitigated by integrating accelerometer-based ANC.

3.5.2. Reasons for 79.2% Accuracy in ML for SCI/MS Prognosis

The 79.2% accuracy in ML-based SCI/MS prognosis, derived from a meta-analysis of eight studies (n = 1800, 560 SCI/MS patients), stems from robust feature engineering and model selection. Key predictors include SSEP latency (reflecting conduction delays due to demyelination) and MEP amplitude (indicating axonal integrity), which are strongly correlated with clinical outcomes (ASIA scale for SCI, EDSS for MS) [2]. Random Forest (RF) and Neural Network (NN) models excel in capturing non-linear relationships, achieving 70–98% accuracy across studies, with NNs reaching 98% for MS progression due to their ability to model complex temporal patterns [22]. Multimodal integration of EPs with MRI (e.g., DTI fractional anisotropy) boosts accuracy by 5–10% by providing structural insights [2]. However, moderate heterogeneity (I² = 56–62%) from varying EP protocols (e.g., stimulus intensity, electrode placement) and retrospective data limits the pooled accuracy to 79.2%. Compared to traditional EP analysis (60–70% accuracy without ML), the use of ensemble methods (RF, SVM) and feature selection (e.g., SHAP-based latency prioritization) improves generalizability [26]. Limitations include small sample sizes in some studies (e.g., n = 32, [24]) and potential bias from retrospective designs, which may underestimate accuracy in prospective settings. The meta-analysis reveals neural networks achieve up to 95.0–99.0% accuracy in SCI/MS prognosis, compared to 70–84.7% for random forests (Figure 3).

3.6. Computational Intelligence in Physiological Measurements

Physiological measurements face variability from biological, environmental, and instrumental sources. Medical physicists apply the Guide to the Expression of Uncertainty in Measurement (GUM) framework, achieving ±2.8 ms precision in ECG RR-intervals [27]. Bayesian inference and Monte Carlo simulations provide robust confidence intervals [28].

In EEG, uncertainty quantification ensures accurate spectral edge frequency calculations, meeting IEC 60601-2-26 standards [29]. In positron emission tomography (PET), statistical models quantify tracer uptake uncertainties [30]. Monte Carlo dropout estimates uncertainty in Neural Networks for EEG classification by applying dropout at inference, generating multiple predictions (Algorithm 3).

For a CNN classifying ictal patterns, the variance of (T) predictions quantifies uncertainty, achieving 5.1% false positives [31].

The dropout probability is as follows:

Algorithm 3: Monte Carlo Dropout

[ p = 0.5 ] Variance (\sigma) is: [ \sigma = \frac{1}{T} \sum_{t = 1}^{T} (y_t − \bar{y})^2 ] Algorithm: Monte Carlo Dropout for EEG Uncertainty Input: EEG data (x), trained CNN (M), dropout rate (p) Output: Prediction (y), uncertainty (\sigma) Initialize empty lists (Y), (V) For (t = 1) to (T): a. Apply dropout to (M) with rate (p) b. Predict (y_t = M(x)) c. Append (y_t) to (Y) Compute mean (y = \text{mean}(Y)) Compute variance (\sigma = \text{var}(Y)) Return (y), (\sigma)

For EPs, uncertainty frameworks validate predictors like SSEP amplitude, correlating with MRI fractional anisotropy ((r \approx 0.3–0.5)) in SCI [2,32]. This rigor supports regulatory compliance and clinical trust.

3.7. Federated Learning for EP-Based ML Models

Computational intelligence enhances physiological measurements by processing complex datasets. Medical physicists deploy 1D CNNs (68 KB quantized) on wearables, achieving 94.2% accuracy in arrhythmia detection with 12 μW power, enabling week-long monitoring [13]. The CNN architecture includes the following (Algorithm 4):

Pseudo-code for 1D CNN for Arrhythmia Detection

Algorithm 4: CNN for Arrhythmia Detection

model = Sequential() model.add(Conv1D(filters = 32, kernel_size = 5, activation = ‘relu’, input_shape = (128, 1))) model.add(MaxPooling1D(pool_size = 2)) model.add(Conv1D(filters = 64, kernel_size = 3, activation = ‘relu’)) model.add(MaxPooling1D(pool_size = 2)) model.add(Flatten()) model.add(Dense(128, activation = ‘relu’)) model.add(Dense(5, activation = ‘softmax’)) # 5 arrhythmia classes model.compile(optimizer = ‘adam’, loss = ‘categorical_crossentropy’, metrics = [‘accuracy’])

In EEG, edge-AI detects ictal patterns with a 5.1% false positive rate versus 8.2% in hospital systems, reducing latency to 8 s [31]. Federated learning enables privacy-preserving multicenter validation:

Pseudo-code for Federated Learning Workflow

def federated_train(local_models, global_model):   global_weights = global_model.get_weights()   for client in clients:    local_model = local_models [client]    local_model.train_on_local_data()    local_weights = local_model.get_weights()    global_weights = aggregate_weights(global_weights, local_weights, client_weight)   global_model.set_weights(global_weights)   return global_model

In precision neurology, ML integrates EPs with imaging, predicting SCI/MS outcomes with 79.2% accuracy [2]. In functional MRI, Deep Learning corrects motion artifacts in real time, enhancing neural signal detection [33].

3.8. Explainable AI for EP-Based Models

Federated learning enables privacy-preserving ML model training across institutions, critical for scaling EP-based prognostic models. Each site trains a local model on EP data (e.g., SSEP latency, MEP amplitude), sharing only model weights with a central server for aggregation. This approach achieves 79.2% accuracy while complying with GDPR and HIPAA [34].

Federated learning’s ethical implications are profound. While it preserves patient privacy by avoiding data centralization, data heterogeneity across sites (e.g., variations in EP protocols or patient demographics) can introduce bias, potentially leading to unfair models that disadvantage underrepresented groups (e.g., low-income populations with limited access to advanced EPs). Ethical strategies include auditing for bias using metrics like the disparate impact ratio and ensuring equitable data representation through weighted aggregation. Informed consent must emphasize federated learning’s benefits and risks, such as model degradation from heterogeneity. Regulatory frameworks like GDPR require transparency in weight aggregation to mitigate risks [34]. These measures ensure ethical deployment of federated learning, enhancing model generalizability for SCI/MS prognosis [2,35].

3.9. Imaging Correlations: Complementing EPs with Structural Insights

Explainable AI (XAI) enhances trust in EP-based ML models. SHAP values quantify feature contributions (e.g., SSEP latency), revealing that latency accounts for 42% of SCI prognosis accuracy (Algorithm 5), [36].

Algorithm 5: SHAP for EP-Based ML

Input: EP data (x), ML model (M) Output: SHAP values (\phi) Initialize background dataset (X_b) For each feature (i) in (x): a. Compute predictions for all coalitions (S \subseteq X_b) b. Calculate (\phi_i = \sum [weight(S) ∗ (M(S \cup i) − M(S))]) Return (\phi)

3.10. Role of Medical Physicists in Electrophysiological Applications

Imaging modalities enhance EP-based prognostics by providing structural insights [32,37]. T2-weighted MRI reveals edema in SCI and demyelinating plaques in MS, correlating with EP abnormalities [24]. Diffusion tensor imaging (DTI) shows reduced fractional anisotropy (FA), indicating axonal disruption, aligning with SSEP amplitude reductions ((r \approx 0.3–0.5)). Magnetization transfer imaging (MTI) detects myelin loss, complementing EP latency delays [9]. Magnetic resonance spectroscopy (MRS) identifies reduced N-acetylaspartate, signaling axonal loss, corresponding to absent EP amplitudes [32]. Medical physicists integrate these modalities with EPs, enhancing ML model accuracy [2,35].

3.11. Quantum Sensors for High-Resolution EEG

Medical physicists are instrumental in electrophysiological applications, spanning ECG, EEG, SSEPs, and MEPs for diagnosing cardiac and neurological disorders [10,38]. They design systems with microelectrode arrays, achieving 8–12 dB SNR improvements in EEG via WPD, enabling 41% more spike–wave complex detection in epilepsy monitoring [19,31]. In neonatal seizure detection, AI models achieve 96.4% sensitivity with 8.3 s latency, surpassing clinical observation [31].

In Parkinson’s disease, wearables monitor β-band (13–30 Hz) activity, achieving 94% burst detection versus 68% in clinical EEG, reducing DBS adjustment lags from weeks to 48 h [39,40]. In cardiac monitoring, edge-AI detects arrhythmias with 94.2% accuracy, reducing false positives by 5–7% [25]. A meta-analysis analyzed eight studies (n = 1800), showing that EP-based ML predicts SCI recovery (ASIA scale) and MS progression (EDSS) with 80.0% accuracy and an AUC of 0.83 [2]. Their quality assurance, including MDS compliance and electromagnetic shielding, ensures reliable diagnostics [2].

4. Meta-Analysis: Methodology and Findings

Medical physicists’ methodological interventions are exemplified by a meta-analysis evaluating EP-based ML models for SCI/MS prognosis [2]. The search strategy involved databases like PubMed, Scopus, and Web of Science, searched up to May 2025 using keywords such as “spinal cord injury,” “multiple sclerosis,” “evoked potentials,” “machine learning,” and “prognosis,” with MeSH terms (e.g., “Evoked Potentials,” “Spinal Cord Injuries,” “Multiple Sclerosis,” “Machine Learning”) and filters for human studies and meta-analyses. The search strategy included Boolean operators (e.g., “evoked potentials AND (spinal cord injury OR multiple sclerosis) AND machine learning”) and manual screening of reference lists for additional studies. Inclusion criteria required studies applying ML to EPs (SSEP/MEP) for predicting SCI recovery (ASIA scale), injury location, or MS progression (EDSS), reporting accuracy/AUC, and using validated ML methods (e.g., cross-validation). Exclusion criteria eliminated studies without ML, non-EP biomarkers (e.g., solely imaging-based), or missing performance metrics. Eight studies (n = 1800) were included, covering SCI (n = 863) and MS (n = 937), using SSEPs and MEPs to predict ASIA recovery, injury location, or EDSS progression. A random-effects model (DerSimonian-Laird) pooled estimates, accounting for heterogeneity (I² = 56–62%). Sensitivity analyses excluded an animal study.

Results: The pooled analysis of eight studies revealed an accuracy of 80.0% (95% CI 77.5–82.5%) and an AUC of 0.83 (95% CI 0.80–0.86). SSEP latency and MEP time series were universal predictors, with amplitude critical for SCI. Multimodal integration (SSEP-MRI) enhanced predictions (

Table 1. Characteristics of studies included in the meta-analysis, including population, EP type, ML model, outcome, accuracy, and AUC.

Study	Population	EP Type	ML Model	Outcome	Accuracy (%)	AUC
[2]	MS, n = 125	SSEP	Random Forest	EDSS progression	75.0 (70.2–79.8)	0.78 (0.73–0.83)
[2]	SCI, n = 123	SSEP	Random Forest	ASIA recovery	83.0 (78.4–87.6)	0.87 (0.83–0.91)
[21]	MS, n = 80	MEP	Random Forest	EDSS progression	70.0 (64.8–75.2)	0.75 (0.70–0.80)
[24]	SCI, n = 32	SSEP	Random Forest	Injury location	84.7 (79.9–89.5)	0.85 (0.80–0.90)
[41]	SCI, n = 80	SSEP	Deep Learning	ASIA recovery	81.2 (76.0–86.4)	0.83 (0.79–0.87)
[23]	MS, n = 740	Clinical/MEP	Neural Network	RR-to-SP conversion	98.0 (95.0–99.0)	0.98 (0.95–0.99)
[6]	MS, n = 500	MRI/EP	CNN	EDSS > 3	85.0 (80.0–90.0)	0.89 (0.85–0.93)

). These findings suggest that ML methods like NN and SVM complement RF and DL, improving generalizability for SCI/MS prognosis. Higher accuracy in SCI studies reflects stronger EP associations with axonal damage, linking to discussion on clinical translation [2].

Insights: Higher accuracy in SCI (81.2–84.7%) versus MS (70–98%) reflects stronger EP associations with axonal damage in SCI, with newer methods like NN achieving higher performance [2]. These results link to the discussion on expanding ML applications for better clinical translation (

Table 2. ML feature importance in EP-based models.

Feature	Importance (Mean ± SD)	Application
SSEP Latency	0.42 ± 0.08	SCI/MS prognosis
MEP Time Series	0.35 ± 0.06	SCI/MS prognosis
SSEP Amplitude	0.18 ± 0.04	SCI prognosis
MRI Fractional Anisotropy	0.05 ± 0.02	SCI prognosis

and

Table 3. Heterogeneity sources.

Source	I2 (%)	Impact
EP Type (SSEP vs. MEP)	58	Moderate
ML Model Type	62	High
Sample Size	56	Moderate

).

Impact on Clinical Outcomes

Medical physicists’ innovations enhance clinical outcomes. Wearable EEG with WPD reduces indeterminate readings by 28%, detecting 41% more spike–wave complexes [31]. Edge-AI enables earlier arrhythmia detection, reducing false positives by 5–7% [25]. In SCI, EP-based ML predicts ASIA score improvements with 83% accuracy, guiding rehabilitation [2]. In MS, models predict EDSS worsening with 70–98% accuracy, informing early interventions [21,23]. In Parkinson’s, β-band monitoring reduces DBS adjustment lags to 48 h [39]. These advancements support personalized medicine and regulatory compliance [28,42]. The 79.2% prognostic accuracy for SCI/MS exceeds traditional clinical assessments, which typically achieve 60–70% accuracy for EP alone [7,8]. For SCI, ML-enhanced EP models (83% accuracy) outperform serum neurofilament light chain (NfL) biomarkers (70% sensitivity) by enabling earlier and more precise rehabilitation planning, such as transcranial stimulation [2]. In MS, ML models (70–98% accuracy) surpass clinical observation (50–60% for EDSS prediction), facilitating timely disease-modifying therapies like ozanimod [21,23]. These improvements reduce diagnostic delays and optimize treatment outcomes (

Table 4. Clinical scenarios illustrating applications of medical physicists’ innovations in wearable electrophysiology and precision neurology.

Scenario	Condition	Description
1	Spinal Cord Injury	A 35-year-old with a C6 incomplete injury undergoes SSEP assessment. An ML model predicts 83% likelihood of ASIA score improvement, guiding intensive physiotherapy and transcranial stimulation [35].
2	Multiple Sclerosis	A 42-year-old with relapsing–remitting MS shows prolonged SSEP latency. An ML model predicts 70% probability of EDSS worsening, prompting early ozanimod initiation [21]
3	Epilepsy	A 27-year-old with focal epilepsy uses a wearable EEG, detecting 41% more spikes, refining surgical planning [31].
4	Cardiac	A 60-year-old with suspected atrial fibrillation uses a wearable ECG with edge-AI, achieving 94.2% accuracy, enabling timely anticoagulation [25].
5	Stroke Monitoring	A 55-year-old post-stroke patient uses wearable EEG with AI, detecting epileptiform activity (80% sensitivity), guiding levetiracetam therapy to prevent seizures [26].
6	Atrial Fibrillation in Elderly	An 83-year-old frail patient uses a wearable ECG for 72 h monitoring, with edge-AI detecting paroxysmal atrial fibrillation (92% specificity), supporting apixaban initiation [25].
7	Chronic Pain in SCI	A 40-year-old with SCI undergoes SSEP testing. An ML model predicts 75% likelihood of neuropathic pain, recommending spinal cord stimulation [43].

).

5. Challenges and Future Directions

Challenges include limited recognition and resource constraints in low-income settings [33]. NV-diamond magnetometers’ power consumption (10–100 mW) limits wearable EEG applications [44]. Strategies to mitigate this include pulsed excitation schemes to reduce microwave power broadening, lowering consumption by 20–30%; compact, low-power laser diodes (e.g., vertical-cavity surface-emitting lasers) replacing bulky lasers, reducing power to 5–10 mW; and optimized pulse sequences for efficient NV center readout, achieving 50% power savings [44,45]. These approaches aim to reduce consumption to 1–10 mW, enabling practical wearable EEG systems. Moderate heterogeneity in ML models (I² = 56–62%) and limited EP studies (n = 8) constrain generalizability [2]. Multicenter validation and standardized EP protocols are needed.

Future research should prioritize federated learning for privacy-preserving data pooling and quantum-enabled sensors for higher resolution [34,44]. Integrating medical physicists into FDA/CE standardization committees will ensure regulatory-compliant devices, addressing barriers like power efficiency and validation gaps [42]. Joint training initiatives, such as those for Monte Carlo dropout uncertainty quantification [31], foster shared understanding of AI limitations (e.g., (\sigma > 5%) indicating low-confidence predictions). Embedding medical physicists into clinical teams, as seen in epilepsy monitoring units [46], ensures that technologies meet bedside needs while adhering to IEC 60601-2-47 safety standards [14]. This collaborative ethos, championed by [28], will accelerate the transition from lab innovation to routine care.

6. Limitations

Several limitations must be acknowledged in these advancements. First, while AI models demonstrate high accuracy (e.g., 79.2% for SCI/MS prognosis), their performance may vary across patient populations due to heterogeneity in EP protocols and disease subtypes, as seen in the meta-analysis by [2] where I² values of 56–62% indicated moderate methodological diversity. Second, quantum sensors like NV-diamond magnetometers, despite their 10 pT/√Hz sensitivity, face practical barriers such as 10–100 mW power consumption [44], limiting wearable applications. Third, federated learning assumes consistent data quality across sites; variations in EP acquisition (e.g., stimulus intensity, electrode placement) could introduce bias, as noted in multicenter EEG studies [31]. Addressing these challenges requires stricter standardization, exemplified by IEC 60601-2-47 guidelines for wearable devices [14].

7. Conclusions

Medical physicists and biophysicists are pivotal in advancing wearable electrophysiology and precision neurology, delivering innovations like 60% motion artifact reduction, 94.2% accurate AI diagnostics, and 79.2% accurate EP-based prognostics for SCI/MS [47]. Their expertise in signal processing, AI, uncertainty quantification, and methodological research ensures reliable, regulatory-compliant technologies. By embedding these professionals in clinical and regulatory frameworks, healthcare systems can enhance diagnostic precision, optimize treatments, and improve patient outcomes, shaping the future of personalized medicine. Bridging disciplinary silos is paramount for translational success [48]. The Movement Disorders Society’s guidelines for wearable sensors [39] demonstrate how standardized workflows can align engineers, clinicians, and physicists. Joint training initiatives, such as those for Monte Carlo dropout uncertainty quantification [31], foster shared understanding of AI limitations (e.g., (\sigma > 5%) indicating low-confidence predictions). Embedding medical physicists in clinical teams, as seen in epilepsy monitoring units [31], ensures that technologies meet bedside needs while adhering to IEC 60601-2-47 safety standards [14]. This collaborative ethos, championed by [28] will accelerate the transition from lab innovation to routine care.