Open research questions in Machine Learning in Healthcare

What the literature leaves open

• While limited by the sample size of the primary cohort preventing the use of deep learning architectures, our study provides a validated, cost-effective pipeline for early COVID-19 screening and triage. For general screening populations (Data4u), appropriate imputation (specifically MICE) can improve sensitivity by up to 26 percentage points, enabling the recovery of critical diagnostic signals from sparse data.

  Mission imputable: Effects of missing data processing on infectious disease detection and prognosis · 2026 · DOI

• FUTURE SCOPE Looking forward, future work should focus on several key areas to expand the system’s usefulness and strength:  Data Expansion: We could enhance capabilities to man-age larger and more varied datasets, including patient histories, genetic data, and real-time health monitoring information.

  A Review of Machine Learning Techniques for Symptom-Based Disease Prediction · 2026 · DOI

• Abstract Objective Long COVID (LC) remains poorly understood, and there is a critical need for advanced computational tools to better identify and characterize patients.

  Characterization and Validation of EHR Computable Phenotypes for Long COVID Using Patient-Reported Symptoms: Insights from the Nationwide RECOVER Program · 2026 · DOI

• The majority of AI models in neuroimaging-based psychiatry carry a high risk of bias, external validation remains rare, and evidence of real-world clinical impact is scarce.

  Artificial intelligence in psychiatry: clinical applications, limitations, and ethical challenges · 2026 · DOI

• Large Language Models (LLMs) Recent work has begun to explore LLMs (e.g., GPT-4, LLaMA) for automated ICD coding primarily in assistive or prompt-based roles, including sentence-level extraction and few-shot or zero-shot coding, rather than as fully end-to-end classifiers. Early studies suggest the potential of prompt-based approaches using pretrained language models for ICD coding assistance; however, widespread adoption remains constrained by data privacy requirements, reliance on proprietary systems, reproducibility challenges, and the absence of standardized evaluation protocols for clinical coding tasks. Multimodal Data Integration An emerging direction in automated ICD coding is the integration of clinical text with structured EHR data, such as laboratory results, vital signs, and diagnosis histories. Combining unstructured notes with structured variables can improve coding for complex admissions and comorbid conditions, particularly when key diagnostic evidence is sparsely documented in text alone. However, reported performance gains are often modest and highly dataset-dependent, and multimodal systems introduce additional challenges related to data heterogeneity, feature alignment, and cross-institution portability, which have limited their widespread adoption in current benchmark-driven research [31, 74]. Explainable AI for Clinical Trust Future research should prioritise clinician-friendly interfaces that provide granular, natural-language explanations of code assignments, quantify model confidence, and highlight critical text segments, going beyond attention visualisation toward methods such as SHAP and LIME that offer more faithful explanations. Regulatory frameworks such as the EU’s AI Act underscore the urgency of this direction, making explainability a pressing design requirement rather than an optional addition. Real-World Deployment and Validation Translating research into practice requires addressing computational constraints, integrating models into EHR systems, and validating performance across diverse healthcare settings (e.g., urban vs. rural, high- vs. low-resource). Pilot studies in realworld environments, coupled with clinician feedback, could bridge the gap between research and application. Ethical and Regulatory Considerations As automated ICD coding systems scale, ethical challenges, data privacy, algorithmic bias, and equitable access must be addressed. Robust governance frameworks, informed by interdisciplinary collaboration (e.g., clinicians, ethicists, policymakers), will ensure responsible deployment.

  Recent Advances in AI for Automated ICD Coding: A Systematic Literature Review · 2026 · DOI

• Obviously, the utility of any data-driven approach is dependent on the data used to train the models, and much future research is needed. Finally, as noted previously, a key limitation is that the curated real dataset underrepresents the diversity and unpredictability typical of everyday practice set- tings.

  Machine Learning to Analyze Alternating Treatments Graphs · 2026 · DOI

• 406 Future research should incorporate other methods to improve extraction 407 accuracy of NAACCR variables, particularly through advanced, targeted prompting 408 strategies, and the integration of retrieval-based methods to better manage long clinical 409 contexts. Additionally, expanding the set of NAACCR variables and investigating results 410 stratified by primary site would provide additional insight into the broader applicability of 411 this work. Researchers could also study whether smaller, clinically focused models 412 might outperform the general-purpose foundation models evaluated in this study while 413 offering lower inference costs. Continued investigation is also required to determine how 414 LLM-based systems can be effectively integrated into clinical data workflows in a 415 manner that supports, rather than replaces, expert human judgment. One such option 416 could be a comparison between LLM-augmented registry abstraction and standard 417 registry abstraction to directly compare time, accuracy, and cost with a goal of 22 medRxiv preprint doi: https://doi.org/10.64898/2026.06.25.26356626 ; this version posted June 29, 2026. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY 4.0 International license . 418 enhancing registry efficiency and timeliness while supporting the accuracy appropriate 419 for the clinical application. 420 This study provides a systematic evaluation of LLM performance on a realistic 421 cancer registry abstraction task and highlights both the potential and current limitations 422 of these models in this domain. Although LLMs are not yet a complete solution for 423 automated registry abstraction, they represent a significant step toward more scalable 424 and efficient clinical data extraction.

  Large language models for cancer registry abstraction: a real-world evaluation across models, variables, and cancer types · 2026 · DOI

• Specialist epilepsy expertise is scarce in resource-constrained settings, making LLM-based decision support attractive for frontline clinicians managing longitudinal treatment.

  Teaching LLMs to Recommend and Defer in Underrepresented Epilepsy Care · 2026

• 1 Main findings Taken together, the results support the central premise: cross- model disagreement is sparse enough to be useful as a triage signal and concentrated enough in meaning-bearing differences to substantially exceed random-flagging rates, though the modest positive predictive value (∼56% at A ≤ 3) means that roughly two in five flagged high-risk tokens were not HC- verified errors on this corpus.

  Cross-model disagreement as a reference-free signal for prioritizing human review in medical speech transcription · 2026 · DOI

• recognition and multi-drug modeling remain. Future development will incorporate advanced fuzzy-matching techniques, tighter provenance enforcement, and formal clinical validation to maintain the AI’s transparency and reliability in real-world medical workflows.

  Reducing cognitive load in polypharmacy: A prototype clinical decision support system · 2026 · DOI

• We found that a small symmetric CRF head was most useful in the diabetes task, especially when training data were limited, while no single interaction head dominated in myocardial infarction.

  Comorbidity structure as an inductive bias: Comparing output-head designs for multi-label prediction of diabetes and myocardial infarction complications · 2026 · DOI

• Future research should investigate semi-supervised continual learning techniques [27] capable of updating the memory banks with pseudo-labels derived from the model’s own predictions in the absence of immediate true labels. Future work should focus on grounding prototypes in clinically defined phenotypes, as well as developing theoretical guarantees for conformal continual learning under non-exchangeable data streams, and addressing scalability via hierarchical prototype retrieval mechanisms.

  Conformal Memory-Augmented Attention Networks for Robust and Adaptive Disease Prediction · 2026 · DOI

• Background: National electronic health record (EHR) networks can support learning health systems (LHSs) by enabling large-scale data aggregation, monitoring, and benchmarking, but their capacity to produce trustworthy and locally deployable machine learning and artificial intelligence (ML/AI) models remains uncertain.

  Opportunities and Challenges in Using National EHR Networks for AI in Learning Health Systems · 2026 · DOI

• Buttock claudication after endovascular aneurysm repair (EVAR) impairs recovery and quality of life, yet individualized preoperative risk tools are scarce.

  Development and Internal Validation of an Explainable Machine Learning Model for Predicting Buttock Claudication After EVAR: A Dual-Center Cohort Study · 2026 · DOI

• In this sparse and irregular ADNI setting, transition-based modelling of adjacent visits achieved higher predictive accuracy than the sequence-based branch, suggesting that local transition modelling may be more data-efficient.

  Transition-Based Digital Twin Modelling for Alzheimer's Disease under Sparse Longitudinal Data · 2026

• The benefit of combining both modalities was most evident at earlier horizons, when physiological data were sparse: agreement between the two specialists dropped by more than half from 48 to 6 hours, while the median contribution from clinical notes increased from 37% to 49%.

  Early-Horizon Multimodal ICU Mortality Prediction Without Retraining · 2026 · DOI

• However, a key limitation of SuStaIn is its assumption of monotonic disease progression, which may be biologically implausible and clinically restrictive, particularly for diseases that commonly involve remission and recovery (e.

  Adapting Monotonic Subtype and Stage Inference (SuStaIn) to Model Patterns of Disease Progression for Psychiatric Disorders · 2026 · DOI

• ↑AUROC; better calibration Workflow effects; fairness limits Legend: NI = non-inferior; ↑/↓ = increase/decrease; Pros = prospective; RCT = randomized trial; Ext = external validation; CAD = computer-aided detection.

  AI Nexus for Early Disease Detection and Risk Prediction: Revolutionizing Healthcare through Intelligence · 2026 · DOI

• Our method has several limitations. First, we interpret the probabilities at the output of the neural classifier as if they are a reflection of its confidence on the presence of the symptom in the text. However, this is not necessarily the case, as neural classifiers are known to have issues with calibration [47]. Still, this is offset by the consistency node, which improves calibration compared to using virtual evi- dence alone, as evidenced by the elevated Brier scores. Second, we make strong assumptions on the types of con- ditional distributions that are learned in the BN. In our case, these assumptions match up perfectly with the true data gen- erating process of the data as described in Rabaey et al. [1]. However, in a realistic setting, one would not have access to the true type of probability distribution for each variable in the network, and would instead need to consult an expert. Third, it can be very challenging to come up with a DAG structure that accurately captures reality. To mitigate this, one could work with a panel of experts who iteratively improve the DAG. Furthermore, future work can focus on using (par- tial) structure learning algorithms [35, 48] to learn the DAG from the data, filling in the gaps where experts are unsure, while still asking experts to validate the final DAG. Note that while the inclusion of the BN in our method might limit its generalization to broader contexts, we explicitly choose to trade in this flexibility for interpretability and expert input. Finally, and related to the previous point, we only vali- dated our method on a single simulated use-case. While this shows the merit of our method as a proof-of-concept and allows for reproducible validation, future work should focus on putting the theory into practice and applying our method to a more realistic and challenging dataset. To this end, the MIMIC-III [49] and MIMIC-IV [50] datasets come to mind.

  Patient-level information extraction by consistent integration of textual and tabular evidence with Bayesian networks · 2026 · DOI

• allowing them to interpret why a system suggests a particular care pathway [35]. This interpretability strengthens trust, as transparent www.ijcat.com 213 International Journal of Computer Applications Technology and Research Volume 12–Issue 12, 202 – 217, 2023, ISSN:-2319–8656 DOI:10.7753/IJCATR1212.1021 nurses can cross-reference machine outputs with clinical judgment before implementing interventions. One promising direction is federated learning, which enables AI models to be trained across multiple healthcare institutions without directly sharing patient data [33]. This method preserves privacy while pooling knowledge from diverse populations, ultimately improving predictive accuracy and reducing the risk of biased recommendations. context-aware Enhanced model architectures now combine predictive analytics with enabling recommendations that account for both structured clinical data integrating and unstructured narrative notes contextual insights, AI-CDSS can better adapt to nuanced patient needs in nursing practice. reasoning, [37]. By Moreover, advances in continuous learning frameworks allow models to update with new evidence or local practice changes without requiring complete retraining [40]. This dynamic adaptability aligns with nursing’s need for up-to-date, evidence-informed decision-making. As illustrated in Figure 5, these innovations are not isolated; they interact within an evolving AI-personalised nursing ecosystem that merges data science with frontline clinical workflows [34]. Collectively, these advancements mark a transition from static, opaque tools toward transparent, adaptable AI solutions that enhance nursing autonomy. 8.2 Integration with IoT and Remote Monitoring The fusion of AI-CDSS with Internet of Things (IoT) technologies is enabling personalised care to extend beyond the hospital into home-based follow-up. IoT devices such as wearable heart monitors, glucose sensors, and smart pill dispensers can transmit continuous streams of patient data to nursing teams for real-time monitoring [38]. of signs AI algorithms process these incoming data to detect early warning proactive deterioration, interventions that may prevent hospital readmissions [36].

  Advancing Patient-Centered Nursing Practices Through AI-Driven Clinical Decision Support Systems and Personalized Care Plans · 2026 · DOI

• behind each As summarised in Table 1, each AI method corresponds to practical nursing applications, from triaging emergency department arrivals to tailoring rehabilitation schedules in physiotherapy wards. These methods not only enhance accuracy but also expand the range of clinical scenarios where decision support can be deployed without overwhelming staff or compromising safety. 3.3 Integration with Nursing Workflows Successful AI-CDSS implementation depends on how seamlessly it integrates into nursing workflows. In most hospital settings, the primary integration point is the EHR platform, where AI-generated alerts and recommendations appear alongside conventional patient charts. This colocation avoids the need for nurses to toggle between multiple systems, reducing cognitive load. For example, in medication administration, AI modules embedded within the EHR can automatically flag potential dosing errors or contraindications before the nurse finalises the order. In acute care units, predictive models monitoring vital signs feed directly into bedside devices, issuing visual and auditory alerts when thresholds are crossed. These alerts are to minimise unnecessary tiered interruptions while ensuring that critical warnings are acted upon immediately. Mobile health (mHealth) platforms represent another vital integration channel. Through secure applications on hospital-issued tablets or smartphones, nurses can receive patient-specific care reminders, review AI-curated educational resources, and update clinical observations in real time. This mobility is particularly valuable in community health nursing, where field visits require rapid access to centralised patient data. www.ijcat.com 205 International Journal of Computer Applications Technology and Research Volume 12–Issue 12, 202 – 217, 2023, ISSN:-2319–8656 DOI:10.7753/IJCATR1212.1021 and clinical Integration also involves aligning AI recommendations with established documentation protocols requirements. In many cases, AI outputs are accompanied by hyperlinks to relevant clinical guidelines, ensuring that decision support aligns with institutional policies and regulatory standards. This compliance-aware design not only increases adoption but also supports audit readiness during inspections. role.

  Advancing Patient-Centered Nursing Practices Through AI-Driven Clinical Decision Support Systems and Personalized Care Plans · 2026 · DOI

• Importantly, the implementation of AI in nursing informatics requires robust change management strategies to ensure clinical staff are confident in interpreting and applying. When properly system-generated integrated, as demonstrated by diverse pilot studies, AI tools can seamlessly complement existing workflows without introducing cognitive overload or excessive alert fatigue. This adaptability has been a decisive factor in sustaining adoption rates across multiple clinical environments and in ensuring positive nurse perceptions of AI-CDSS utility. 2.3 Evidence for Personalised Care Plans The evidence base for AI-enabled personalised care plans is growing, with multiple studies highlighting improvements in both clinical outcomes and patient satisfaction when CDSS tools are tailored to individual needs. Unlike generic protocols, personalised care plans leverage AI to account for a patient’s comorbidities, lifestyle factors, and historical treatment thereby producing more nuanced responses, recommendations. In cardiovascular care, AI-CDSS platforms have been shown to optimise medication regimens by balancing clinical efficacy with patient-reported side effects. Similarly, in oncology nursing, predictive algorithms can anticipate a patient’s tolerance to specific chemotherapy cycles, enabling supportive care adjustments that improve adherence rates. These systems also provide dynamic updates to care plans as new clinical data become available, ensuring that interventions remain aligned with a patient’s evolving condition. incorporate communication modules From a patient engagement perspective, AI-personalised plans often that deliver targeted education materials and self-management prompts. These resources are automatically adjusted based on a patient’s health literacy level, cultural background, and preferred communication channels, to higher satisfaction and better adherence to prescribed regimens. leading Evidence also indicates that such systems can significantly reduce hospital readmissions. For example, in patients with heart failure, AI-generated discharge instructions personalised to the patient’s biomarker trends and home environment have resulted in measurable declines in 30-day readmission rates. pathways As illustrated in Figure 1, the integration of AI into the care planning process represents a logical culmination of decades of technological evolution in CDSS. The transition from rigid, standardised personalised to recommendations reflects a broader movement in healthcare toward precision nursing.

  Advancing Patient-Centered Nursing Practices Through AI-Driven Clinical Decision Support Systems and Personalized Care Plans · 2026 · DOI

• Future research should focus on probabilistic calibration and uncertainty estimation techniques, including Bayesian LLM architectures, to improve reliability and confidence estimation. Future research should focus on standardized evaluation frameworks, robust data governance mechanisms, and improved support for low-resource languages to achieve the vision of globally accessible and clinically validated AI-driven healthcare XIX.

  Comprehensive Survey of Retrieval-Augmented, Knowledge-Graph, and Multimodal Large Language Models for Inclusive Healthcare Guidance: Architectures, Benchmarks, and Clinical Deployment · 2026 · DOI

• Future work will focus on integrating multi- source heterogeneous data, enhancing the model’s long-term multi- step prediction capability, expanding the experimental scope to diverse pediatric viral datasets for comprehensive generalization validation, and developing a real-time prediction system for clinical applications. Nevertheless, this study validates the model on only two datasets, and its broader generalization across more diverse viral datasets and clinical scenarios remains to be verified.

  Dynamic temporal partitioning enhanced transformer for pediatric viral load forecasting · 2026 · DOI

• Future work should focus on designing validation studies that meet regulatory standards for safety and effectiveness. Future work will focus on explainable artificial intelligence to build clinician trust, federated learning to combine data across hospitals without violating privacy, contextual awareness to reduce false alarms, edge computing for faster and more private processing, long term trend analysis for chronic disease detection, and regulatory validation through clinical trials.

  Intelligent Healthcare Tracking System Using Predictive Analytics · 2026 · DOI

Most-cited papers in Machine Learning in Healthcare

• Evaluation of clinical prediction models (part 1): from development to external validation · BMJ · 2024 · 510 citations
• Developing clinical prediction models: a step-by-step guide · BMJ · 2024 · 336 citations
• Towards Generalist Biomedical AI · NEJM AI · 2024 · 333 citations
• ChatGPT for shaping the future of dentistry: the potential of multi-modal large language model · International Journal of Oral Science · 2023 · 287 citations
• Evaluation of clinical prediction models (part 2): how to undertake an external validation study · BMJ · 2024 · 235 citations
• AI-Driven Clinical Decision Support Systems: An Ongoing Pursuit of Potential · Cureus · 2024 · 178 citations
• An Explainable AI Paradigm for Alzheimer’s Diagnosis Using Deep Transfer Learning · Diagnostics · 2024 · 126 citations
• Privacy-preserving large language models for structured medical information retrieval · npj Digital Medicine · 2024 · 101 citations
• Foresight—a generative pretrained transformer for modelling of patient timelines using electronic health records: a retrospective modelling study · The Lancet Digital Health · 2024 · 89 citations
• BERT-Based Medical Chatbot: Enhancing Healthcare Communication through Natural Language Understanding · Exploratory Research in Clinical and Social Pharmacy · 2024 · 78 citations

Most recent work

• Agentic AI in Healthcare: Opportunities, Challenges, and Future Directions · ACM Computing Surveys · 2026
• Drug-drug interaction identification using large language models · medRxiv · 2026
• Explainability in action: A metric-driven assessment of local explanations for healthcare tabular models · medRxiv · 2026
• Architecture and Design of a Prototype Multi-Modal Clinical Decision Support System for Integrative Medicine · Longevity Horizon · 2026
• Opportunities and Challenges in Using National EHR Networks for AI in Learning Health Systems · Learning Health Systems · 2026
• High‐Fidelity Synthetic Data Replicates Clinical Prediction Performance in a Million‐Patient Diabetes Cohort · Advanced Science · 2026
• Emerging opportunities and challenges for clinical prediction models in psychiatry: commentary, Oliver et al · The British Journal of Psychiatry · 2026
• FDS-CAP: Modeling Fragmented Disease Subgraphs with Component-Level Attention for Comorbidity Prediction · Journal of Computational Biology · 2026
• Representation learning to advance multi-institutional studies with electronic health record data from US and France · Nature Communications · 2026
• A domain-adapted large language model to support clinicians in psychiatric clinical practice · Nature Machine Intelligence · 2026

Related topics in Computer Science