Exploring Computational Methods to Advance Clinical Decision Making

The papers comprising our Special Issue on advanced decision making in clinical medicine (which is part of the series in applied bioscience and bioengineering) were conceived and written during a most challenging time for medical science and its delivery—the COVID-19 pandemic. Particularly in the pre-vaccine period, terms and concepts which are normally reserved for specialist journals were, for the first time, the parlance of the daily news. Public health officials and practitioners were called on to explain their technical frameworks and jargon to a wide audience. For most, it was a revelation to learn of the highly mathematical nature of modern medical practice. For many, this was reassuring, and for others, the message was foreign and fueled distrust.

Recent developments in artificial intelligence, data science, and statistics, however, have the potential to enhance the quality of medical decision making and bolster public confidence in the value of the offered solutions. There is a communication gap that is as much the responsibility of scientists and practitioners as it is of the media and politicians.

The landscape of clinical decision making is transforming, driven by the convergence of data science, systems engineering, cognitive analytics, and machine learning. As the complexity and volume of patient data expand thanks to genomic, imaging, sensor-derived, and electronic health record sources, the demand for robust, explicable, and adaptive decision-support tools has intensified. The challenge to combine, for the best effect, the knowledge of healthcare professionals (both frontline and their scientist and technical colleagues) with patient preferences, values and risk tolerance, has never been more pressing.

The role of decision theory and probabilistic reasoning has re-emerged in clinical contexts. Shared decision making frameworks, supported by Bayesian models and utilities elicited from patients, are facilitating more personalized and ethically grounded treatment choices [1]. Simulation-based methods have also gained traction. Studies apply agent-based models, system dynamics, and discrete-event simulations to optimize clinical pathways and resource allocation, especially in critical care and pandemic response scenarios [2].

Decision support also requires understanding causality. Advances in causal inference from observational data—through frameworks such as counterfactual reasoning [3], targeted maximum likelihood estimation (TMLE), and instrumental variable methods—enable clinicians to assess treatment effects more robustly in the absence of randomized trials [4].

Recent research has progressed to embed artificial intelligence (AI) in diagnostic and prognostic workflows. For example, ref. [5] demonstrated the potential of deep learning to match or exceed dermatologist-level accuracy in skin cancer classification. The study [6] advocated for a human-centered AI approach that enhances clinician judgment rather than replacing it. At the same time, explainable AI (XAI) approaches, as outlined in [7], have emerged as critical tools to improve transparency and trust in clinical environments, particularly in high-stakes decisions.

Another trend is the use of reinforcement learning (RL) as a decision-theoretic framework for sequential treatment planning, especially in chronic and critical care. The study referenced in [8] demonstrated how off-policy RL could recommend vasopressor and fluid management strategies for septic patients in the ICU, often aligning with expert decisions. Building on this, ref. [9] highlighted the challenges of implementing RL in healthcare, considering aspects of safety, fairness, and interpretability.

In this Special Issue we present state-of-the-art contributions that examine how artificial intelligence, data science, and statistics can improve and enhance effective medical decision making and evidence-based medicine. We curated a collection of studies that address innovative procedures in evidence-based medicine, data analytics advances and risk management in medical treatment decisions, applications of artificial intelligence methods, modeling and simulation in medical decision making, as well as policy development for better evidence-based practices and services. Collectively, these works exemplify the translational potential of advanced decision methodologies to deliver safer, more equitable, and evidence-informed clinical care. We strongly encouraged articles from interdisciplinary teams that include medical professionals, researchers, clinicians, data scientists, and AI experts.

There are eleven state-of-the-art research articles in this Special Issue that are pertinent to the issues covered. Each paper addresses unique aspects of the enhancement of medical decision making through intelligent methods and data analysis. The contributions made by these studies reflect the variety and richness of current research, focusing on topics ranging from improved diagnostics using AI, through fuzzy data analysis for improved medical treatment, to exploration of adverse drug reactions and improved personalized health case and decision making. In the following paragraphs, we shortly describe the articles included in this Special Issue.

Early or timely detection of serious medical conditions is an area where machine learning can considerably improve efficiency of medical care and patient outcomes. The work of Naeem Ullah, Javed Ali Khan, Mohammad Sohail Khan, Wahab Khan, Izaz Hassan, Marwa Obayya, Noha Negm and Ahmed S. Salama, titled An Effective Approach to Detect and Identify Brain Tumors Using Transfer Learning [10], explores the use of pre-trained deep transfer learning (TL) for the detection and recognition of the three types of brain tumors—gliomas, meningiomas, and pituitary tumors. The latter are among the most critical, widespread, and life-threatening illnesses worldwide. More specifically, they assess the performance of nine pre-trained TL classifiers by automatically identifying and classifying brain tumors using a detailed classification method. The TL algorithms are tested on a baseline brain tumor classification (MRI) dataset, which is freely accessible on Kaggle. The deep learning (DL) models are fine-tuned using their default parameters. The authors find that the inceptionresnetv2 TL algorithm yields the best performance and achieves the highest accuracy in detecting and classifying glioma, meningioma, and pituitary brain tumors, thereby ranking as the top classification algorithm, surpassing the other DL algorithms. The authors also verify their results through comparison with hybrid methods, where they employ convolutional neural networks (CNNs) for deep feature extraction and a support vector machine (SVM) for classification.

Another study that deals with improved detection of serious medical conditions is by Andressa C. M. da Silveira, Álvaro Sobrinho, Leandro Dias da Silva, Evandro de Barros Costa, Maria Eliete Pinheiro and Angelo Perkusich, titled Exploring Early Prediction of Chronic Kidney Disease Using Machine Learning Algorithms for Small and Imbalanced Datasets [11]. The authors discuss chronic kidney disease (CKD), which is recognized as a global public health challenge, and is typically identified during the later stages of the condition due to imbalanced and small datasets. The work uses medical record data about Brazilians with and without a CKD diagnosis, which includes attributes such as hypertension, diabetes mellitus, creatinine levels, urea, albuminuria, age, gender, and glomerular filtration rate. They then apply oversampling methods with both manual and automated augmentation techniques, such as the synthetic minority oversampling technique (SMOTE), borderline-SMOTE, and borderline-SMOTE support vector machine (borderline-SMOTE SVM). Modeling is performed using decision trees (DTs), random forests, and multi-class AdaBoosted DT algorithms. The authors also apply methods for dynamic classifier selection, such as overall local accuracy and local class accuracy. For dynamic ensemble selection, they use k-nearest oracles-union, k-nearest oracles-eliminate, and META-DES. The performance of the models is assessed using hold-out validation, multiple stratified cross-validation (CV), and nested CV. The authors show superior accuracy for the decision tree model accuracy through manual augmentation and SMOTE. The outcomes of this work can contribute to the development of systems aimed at the early detection of CKD, particularly when dealing with imbalanced and limited datasets.

A third related work in this Special Issue addresses the use of artificial intelligence (AI) in the healthcare sector—the work of Saleem Ameen, Ming-Chao Wong, Kwang-Chien Yee and Paul Turner, titled AI and Clinical Decision Making: The Limitations and Risks of Computational Reductionism in Bowel Cancer Screening [12]. While AI techniques are often praised as enhancers of precision, safety, and quality of clinical decisions, treatments, and patient care, they depend on reductive reasoning and computational determinism that incorporate problematic assumptions regarding clinical decision making and practice. They tend to simplify the autonomy, expertise, and judgment of clinicians to inputs and outputs that are framed as binary or multi-class classification challenges measured against a clinician’s ability to identify or predict disease conditions. The authors investigate this reductive reasoning within AI systems for colorectal cancer (CRC) to underscore their limitations and dangers. Those refer to the issues caused by intrinsic biases found in retrospective training datasets and the embedded assumptions present in fundamental AI architectures and algorithms. They also relate to the inadequate and limited evaluations performed on AI systems before their integration into clinical practice. Another limitation is the underrepresentation of socio-technical factors concerning the context-specific interactions between clinicians, their patients, and the wider healthcare system. As a result, the authors recommend that to maximize the advantages of AI systems and prevent adverse unintended effects on clinical decision making and patient care, it is essential to adopt more nuanced and balanced approaches to the deployment and evaluation of AI systems in CRC.

Two of the works in the Special Issue explore diagnostics techniques of adverse drug reactions.

The first work is by Jianxiang Wei, Lu Cheng, Pu Han, Yunxia Zhu and Weidong Huang, titled Decision Tree-Based Data Stratification Method for the Minimization of the Masking Effect in Adverse Drug Reaction Signal Detection [13]. They posit that data masking is an inherent flaw in the measures of disproportionality used for detecting signals in adverse drug reactions (ADRs). They introduce a decision tree stratification approach to reduce the masking effect by incorporating both patient- and drug-related factors. They utilize adverse drug reaction monitoring records from the Jiangsu Province in China from 2011 to 2018. The authors define the age divisions for antibiotic-related data and perform correlation analysis on the gender and age of patients in relation to the properties of drug categories. They then develop a decision tree using the J48 algorithm, which classified whether drugs were categorized as antibiotics based on age and gender. They also introduce performance evaluation metrics such as recall, precision, and F score (the harmonic mean of recall and precision). Using four experiments (based on the proportional reporting ratio methodology: non-stratification, gender-stratification, age-stratification, and decision tree stratification), the authors show that decision tree stratification outperformed the other three approaches, and the data-masking effect can be further reduced by thoroughly considering confounding factors related to both patients and drugs.

The second work is by Jianxiang Wei, Jimin Dai, Yingya Zhao, Pu Han, Yunxia Zhu and Weidong Huang, titled Application of Association Rules Analysis in Mining Adverse Drug Reaction Signals [14]. The authors emphasize the use of spontaneous reporting systems (SRSs) as a key method for tracking ADRs that occur during clinical drug use. To detect signals for ADRs, researchers often use disproportionality analysis (DPA) and do not incorporate data mining techniques. Here, the authors rely again on the spontaneous reports from Jiangsu Province, China for the period from 2011 to 2018 and apply association rules analysis (ARA) to extract signals. For their analysis, they define crucial metrics for ARA, e.g., confidence and lift, and develop performance evaluation indicators like precision, recall, and F1 as objective benchmarks. The results show improvement of the F1 score using the ARA method, representing a significant enhancement. To mitigate drug risks and support decision making regarding drug safety, it is essential to integrate and utilize more data mining techniques for ADR signal detection.

The study by Kiril Tenekedjiev, Daniela Panayotova, Mohamed Daboos, Snejana Ivanova, Mark Symes, Plamen Panayotov and Natalia Nikolova, titled Quasi-Experimental Design for Medical Studies with the Method of the Fuzzy Pseudo-Control Group [15], explores the effect of interventions over a given parameter representing the medical condition of patients. The authors propose a novel fuzzy quasi-experimental computational approach, called the method of the fuzzy pseudo-control group (MFPCG), which addresses the limitations of methods commonly used (e.g., the difference-in-differences (DID) method). The method uses four fuzzy samples as input and statistically compares the favorability of the differences in a continuous parameter before and after the intervention for the experimental and the pseudo-control groups. MFPCG applies four modifications of fuzzy Bootstrap procedures to perform each of the nine statistical tests. As a case study, the team explores a dataset related to the effect of annuloplasty that acts in conjunction with revascularization over two continuous parameters that characterize the condition of patients with ischemic heart disease complicated by moderate and moderate-to-severe ischemic mitral regurgitation. The statistical results proved the favorable effect of annuloplasty on two parameters, both for patients with a relatively preserved medical state and patients with a relatively deteriorated medical state. The results of the MFPCG are compared with those of a fuzzy DID. The work discusses the limitations and adaptability of MFPCG, indicating that it is not a technique competing with DID but instead should be used alongside it.

Artificial intelligence, machine learning, and data science are becoming commonplace in data analysis and statistics. However, a versatile software tool tailored to the interactions in small biomedical teams is often missing. Proprietary commercial software packages bear inherent risks of abandonment or unpredictable company policies. Open-source software libraries may require too much effort to use. In the work, titled Anatomy of a Data Science Software Toolkit That Uses Machine Learning to Aid ‘Bench-to-Bedside’ Medical Research—With Essential Concepts of Data Mining and Analysis Explained, the authors László Beinrohr, Eszter Kail, Péter Piros, Erzsébet Tóth, Rita Fleiner and Krasimir Kolev [16] describe a toolkit to address this problem. Their toolkit is designed from bottom to top with small teams in mind, which allows individuals with very different expertise to work together with only specific parts of the software dedicated to their expertise. The proposed approach is based on open-source components (existing modular Python platform libraries); thus, company policy risks are alleviated. This paper also summarizes basic concepts in data science that serve the structured data organization through a contemporary data analysis applied in the described toolkit. The authors also show examples from their laboratory using blood sample and blood clot data from thrombosis patients (suffering from stroke, heart and peripheral thrombosis disease) and how such tools can help to set realistic expectations and show caveats.

A different scope of exploration is offered by Andrzej Walczak, Paweł Moszczyński and Paweł Krzesiński, in their work titled Evolution of Hemodynamic Parameters Simulated by Means of Diffusion Models [17]. They explore the similarity between diffusion as a physical concept in particle movement, heat transfer, etc., and the behavior of medical data. They posit that changes in medical parameters recorded during patient treatment can also be described using diffusion models. They view a patient medical condition by a set of discrete values, and the progression of condition is represented as a transition of continuously changing, temporal attributes from one discrete parameter value to another, linked to given parameters. The capacity to forecast such diffusion-related characteristics provides invaluable support in diagnostic decision making. The authors assess several hundred patients to study how to stabilize their hemodynamic parameters and introduce a diffusion model based on the simulation of treatment outcomes. As a case study, they explore the time evolution of thoracic fluid content (TFC). They use the Fokker–Planck equation (FPE) to verify that the diffusion phenomenon effectively accounts for the changes observed in heart disease parameters.

The study by Joana Magalhães, Maria José Correia, Raquel M. Silva, Ana Cristina Esteves, Artur Alves and Ana Sofia Duarte, titled Molecular Techniques and Target Selection for the Identification of Candida spp. in Oral Samples [18], explores candidiasis and the ability to avoid overuse of antifungal medications. The use of high-throughput technologies for diagnosing yeast pathogens offers notable advantages in terms of sensitivity, precision, and rapidity. While molecular techniques are the subject of considerable investigation, their implementation in clinical practice faces significant obstacles. In their review, the authors explore the pros and cons of existing molecular techniques used for identifying Candida spp., particularly in the context of oral candidiasis. A discussion on their use for diagnosing oral infections seeks to pinpoint the most rapid, cost-effective, precise, and user-friendly molecular approaches suitable for point-of-care testing. The authors pay attention to the challenges healthcare professionals must overcome to ensure an accurate diagnosis.

Another review by Daniela Gifu, titled Soft Sets Extensions: Innovating Healthcare Claims Analysis [19], focuses on the development and use of soft sets and their various extensions and how they are used in the analysis of healthcare claims data. They offer adaptable frameworks for handling the uncertainty and indeterminacy that are characteristic of healthcare claims data. The review traces the evolution of these mathematical tools and how they have advanced healthcare research and improved data analysis techniques. Through real-world illustrations, the author highlights the impact of these tools, emphasizing their critical role in supporting informed decision making and facilitating knowledge discovery within the healthcare sector. The author discussed several case studies to demonstrate the practical usefulness of soft set extensions. The recommendations of this work are to suggest incorporating soft sets and their extensions as a way to enhance the accuracy and effectiveness of healthcare data analysis, leading to improved healthcare results.

In our Special Issue, we were able to accommodate a perspectives paper by Franco Musio, titled The Critical Link in the Successful Application of Advanced Clinical Decision Making—Revisiting the Physician–Patient Relationship from a Practical and Pragmatic Perspective [20]. The author explores the rapidly growing field of advanced clinical decision making influenced by healthcare technologies and the new dimensions of how physicians and patients interact to support medical treatment and shared decision making. The discussion examines the models being researched and utilized in clinical decision making, as well as the connections between physicians and patients, highlighting their complex interactions. Moreover, both clinical decision making and the physician–patient relationship exhibit dynamic, reciprocal relationships that work together in a synergistic manner. The author presents innovative frameworks to clarify these intricate processes, alongside real-life clinical examples. The author widely discusses that the physician–patient relationship serves as a “filter” through which decision-making processes must navigate in order to be executed.

We express our appreciation to the authors for their significant contributions and to the reviewers for their thorough feedback, which allowed us to prepare a Special Issue of such quality. We hope that this Special Issue will motivate future research and findings in innovative data analytics for improved medical decision making, medical diagnostics, and healthcare.