healthcare ml model development

This study addresses the critical yet underexamined role of data filtering in clinical machine learning, which alters statistical structures and directly impacts task complexity and model performance. Despite these effects, existing research frequently treats filtering as routine preprocessing with insufficient transparency. This work reconceptualizes data filtering as a core component of the scientific method, advocating its integration into the broader research paradigm rather than its treatment as a mere technical step. To this end, we develop a transparent and interpretable clinical data preprocessing pipeline and release the corresponding code as open source. Our analysis elucidates the mechanisms through which filtering decisions critically influence data distributions and downstream model efficacy. Ultimately, this research provides a novel framework for enhancing methodological rigor and reproducibility in clinical artificial intelligence studies.

Medical AI deployment is hindered by insufficient production readiness of machine learning (ML) training pipelines. Method: This paper presents a progressive architectural evolution path—monolithic (chaotic) → modular monolithic → microservices—using SPIRA, a voice-based pre-diagnostic system for respiratory insufficiency, as a case study. It systematically introduces continuous training (CT) and a software-quality-attribute-driven MLOps governance framework tailored to healthcare, integrating modular design, microservice decomposition, and engineered CI/CD pipelines. Contribution/Results: The approach significantly improves pipeline maintainability, fault tolerance, and scalability, enabling stable, iterative evolution of SPIRA. It establishes an “agile ML + robust software engineering” co-design paradigm, delivering a reusable methodology and practical benchmark for engineering medical AI in highly regulated environments.

The opacity of black-box AI models in medical devices undermines clinical trust and poses regulatory compliance risks due to insufficient explainability. Method: This study pioneers a systematic integration of eXplainable Artificial Intelligence (XAI) techniques, empirical clinical human–AI interaction research, and regulatory requirements—yielding an XAI integration pathway and tiered validation framework tailored to real-world clinical workflows. Leveraging multimodal expert consensus, in situ clinical behavioral observation, AI explanation evaluation, and MHRA regulatory alignment analysis, we developed a comprehensive XAI implementation guideline spanning development, verification, and deployment phases. Contribution/Results: The guideline was formally adopted by the UK’s Medicines and Healthcare products Regulatory Agency (MHRA) as an official regulatory reference, demonstrably enhancing clinical acceptance and risk controllability. It establishes a methodological paradigm and practical standard for deploying trustworthy, clinically viable AI in healthcare.

Clinical workflow fragmentation severely impedes efficiency: heterogeneous scripting, ad-hoc model ensembles, and lack of data-driven modality identification and standardized outputs result in high deployment overhead, costly monitoring, and poor interoperability. To address this, we propose a healthcare-first vision-language unified framework that pioneers the use of a single vision-language model (VLM) for two-tier clinical decision-making—first, an auditable, three-stage routing mechanism matches inputs to expert-defined model cards; second, domain-specific multi-task joint inference (with early-exit capability and candidate arbitration) adheres to clinical risk constraints. Leveraging phased prompting, a candidate answer selector, and specialty-specific fine-tuning, our framework unifies modality identification, abnormality classification, model selection, and multi-task reasoning. Evaluated across gastroenterology, hematology, ophthalmology, and pathology, our single-model solution achieves performance on par with specialized models while substantially reducing deployment complexity, operational overhead, and integration effort.

Synthetic electronic health record (EHR) data often exhibit insufficient fairness in downstream clinical prediction tasks, undermining equitable AI deployment in healthcare. Method: We propose a task-agnostic, fairness-customizable synthetic EHR generation framework based on conditional generative adversarial networks (cGANs). It jointly models the true EHR distribution and user-specified group fairness constraints—e.g., statistical parity or equalized odds—via fairness-aware regularization to co-optimize fidelity and fairness. Contribution/Results: This work introduces the first synthetic paradigm enabling configurable fairness objectives and decoupling the generator from downstream tasks, addressing the poor generalizability of existing fairness methods in health AI. Evaluated on two real-world EHR datasets across multiple clinical prediction tasks, our method reduces average fairness gaps (ΔDP/ΔEO) by up to 62% while incurring minimal AUC degradation (<1.2%), thus achieving a strong balance between fairness and clinical utility.

This work addresses the mismatch between conventional machine learning practices and the specific performance requirements of clinical tasks in healthcare settings. Traditional approaches rely on differentiable validation losses for model optimization, which often fail to align with clinically meaningful outcomes. To bridge this gap, the paper proposes replacing standard loss functions with non-differentiable yet clinically interpretable custom metrics to guide critical optimization decisions—such as hyperparameter selection and training termination—thereby redefining the model validation pipeline. In two controlled experiments, models optimized using this framework demonstrated significantly superior performance on key clinical tasks compared to those guided by conventional differentiable validation losses. This approach overcomes the inherent limitation of relying solely on differentiable objectives and better aligns medical AI development with real-world clinical goals.

Current machine learning research in surgical risk prediction is often hindered by methodological fragmentation, poor reproducibility, and limited clinical applicability. This study conducts a scoping review of 190 end-to-end machine learning pipelines based on electronic health records, systematically examining critical components including data preprocessing, model selection, evaluation strategies, and interpretability. It presents the first structured synthesis of the entire workflow for surgical risk stratification, uncovering systemic gaps in the use of open datasets, standardized evaluation benchmarks, and deep learning methodologies. The majority of studies rely on single-center proprietary data, and only about one-third incorporate interpretability techniques—factors that severely constrain model generalizability and clinical translation. This work establishes a methodological framework and practical guidance for developing reproducible, generalizable, and clinically viable surgical prediction models.

Prior clinical prediction studies for heart failure lack systematic ablation analyses of preprocessing pipelines and model architectures. Method: Leveraging a large-scale Swedish EHR cohort, we conduct the first comprehensive ablation study in clinical forecasting—systematically evaluating input tokenization strategies, temporal preprocessing techniques, and architectural choices across six sequential models (including Transformer, Llama-enhanced Transformer++, and Mamba), while integrating multimodal time-series data (diagnoses, vital signs, laboratory tests, medications, and procedures). Results: Llama-enhanced Transformer++ and Mamba achieve superior performance over standard Transformer—despite substantially fewer parameters—demonstrating higher accuracy, better calibration, and greater robustness. Notably, both models surpass the large Transformer’s performance using only 75% of the training data. This work establishes a reproducible benchmark for EHR sequence modeling and provides empirically grounded design principles for lightweight, efficient clinical prediction models.

Current medical AI evaluation benchmarks predominantly emphasize knowledge acquisition, failing to adequately capture model reliability, safety, and clinical utility in real-world settings. To address this gap, this work proposes the first systematic evaluation framework aligned with clinical workflows, encompassing end-to-end tasks such as clinical documentation, decision support, and administrative processes. The framework integrates authentic multimodal clinical data and introduces task-specific metrics to comprehensively assess generative models, multimodal systems, and AI agents. Empirical results reveal a substantial performance gap between state-of-the-art models on real-world tasks and their scores on medical knowledge exams—scoring 0.74–0.85 in documentation, 0.61–0.76 in clinical decision-making, and 0.53–0.63 in administrative tasks—highlighting the limitations of existing evaluation paradigms and underscoring the critical role of this framework in advancing the clinical deployment of medical AI.

This work addresses the common lack of continuous evaluation and governance mechanisms in deployed clinical AI systems, which hinders dynamic performance optimization. The authors propose the first end-to-end continuous governance framework tailored for clinical AI, integrating standards-driven validation, A/B testing for controlled version updates, real-time performance monitoring, fault tolerance, and deep integration with electronic health records (EHRs) to establish a closed-loop synergy between engineering iteration and clinical feedback. Applied to Hyperscribe—a speech-to-structured-clinical-note system—the framework achieved substantial improvements over seven iterative cycles: median clinician rating increased from 84% to 95%, negative user feedback decreased from 79% to 30%, median audio processing latency was 8.1 seconds, and task completion rate reached 99.6%, collectively enhancing system reliability and user satisfaction.