Open research questions in Artificial Intelligence in Healthcare

What the literature leaves open

• However, several challenges persist, including data heterogeneity and the lack of standardized protocols, which introduces bias and hinders reproducibility and comparability across studies, as well as concerns about the validity of inference when applying machine learning methods to identify associations between exposure and health outcomes.

  Exposomics and Cardiovascular Diseases: A Scoping Review of Machine Learning Approaches · 2026 · DOI

• Because the model relies only on routinely collected, low-cost variables and open-source software, it is readily transferable to resource-limited settings; future work will focus on prospective, multicentre external validation and on embedding the nomogram into electronic-health-record decision support.

  Explainable XGBoost model and nomogram for risk factor identification and risk prediction in cerebral small vessel disease: a machine learning-based retrospective cohort study · 2026 · DOI

• This model may support individualized long-term risk stratification in patients with concurrent hypertension, HFpEF, and UAP, although prospective validation and clinical impact studies are warranted before routine implementation.

  Risk prediction for long-term cardiovascular events in patients with concurrent hypertension, HFpEF, and unstable angina: a multicenter machine learning-assisted cohort study · 2026 · DOI

• Frontline Medical Sciences and Pharmaceutical Journal 49 FRONTLINE JOURNALS Future research should focus on incorporating multimodal healthcare data, including clinical narratives, laboratory time- series, genomic information, and patient-generated health data, to develop more robust and personalized prediction models.

  Machine Learning-Based Early Prediction of Hospital Readmission Risk Among Chronic Disease Patients Using Electronic Health Records: A Comparative Study of Ensemble Learning Models · 2026 · DOI

• Background Venous thromboembolism (VTE) frequently complicates septic shock, yet precise, individualized risk stratification tools remain scarce.

  Explainable machine learning for predicting venous thromboembolism in septic shock patients · 2026 · DOI

• We describe three alternative calibration approaches when calibration data are lacking: similarity-binning averaging (SBA), adaptive calibration of predictions (ACP), and Elkan calibration.

  Calibrating machine learning approaches for probability estimation without calibration data · 2026 · DOI

• This is particularly relevant as recent reviews of AI- enabled decision support in surgery emphasize explainabil- ity as a key requirement for clinical adoption, which remains insufficiently addressed in many existing models [9, 16].

  Towards clinically interpretable machine learning in emergency surgery: feature importance and insights across clinical time points in abdominal pain cases · 2026 · DOI

• Nevertheless, the main limitation of this study is the relatively limited number of animals, as the dataset was based on 50 Romanov lambs raised under specific farm conditions. Although the high R² values obtained in this study are promising, the results should be validated using larger and independent datasets from different flocks and production systems.

  Prediction of Live Weight in Romanov Lambs Using Body Measurements and Machine Learning Algorithms · 2026 · DOI

• Thus, this future work should focus on privacy preserving AI through Federated learning, model updates, bias aware AI and explainable AI. One major limitation of this study is the usage of a single benchmark dataset namely Cardiotocography (CTG) data set obtained from UCI machine learning repository (Dua and Graff, 2019). Although we tested several classical machine learning, ensemble learning, and neural network based models, the advanced neural network based models like CNNs, RNNs, LSTMs and hybrid models have not been tested extensively due to the lack of more data set in the case of structured and fewer features of the dataset.

  Machine learning-based fetal health prediction and development of smart web application · 2026 · DOI

• Funding Future studies should therefore focus on prospective multicen- ter validation of this model, ideally in larger and more diverse influenza populations. It would also be valuable to examine whether integrating dynamic laboratory trends, radiologic features, and fungal biomarkers could further improve discrimination and cali- bration. Ultimately, a clinically useful IAPA prediction tool should not replace diagnostic judgment, but rather assist clinicians in identifying which influenza A patients require intensified fungal surveillance and earlier diagnostic intervention. The author(s) declared that financial support was received for this work and/or its publication. This work was financially sup- ported by the National Natural Science Foundation of China (Grant No. 82204720); Jiangsu Pharmaceutical Association 2025 Yaoyan Xinsheng Pharmaceutical Research Project (Grant No. 202564123); Jiangsu Society of Research Hospitals Precision Medication - CSPC Pharmaceutical Group Special Research Fund Project (SYHKJ-JY- 2025-25). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

  Machine learning for early screening of influenza A-associated invasive pulmonary aspergillosis in hospitalized patients: a real-world study · 2026 · DOI

• While we note that these findings are important, especially for the identification of certain risk factors and with respect to broadening the population for which POD can be predicted, future work could investigate how our approach generalizes to other datasets in different geographical and ethnic settings.

  Prediction of post-operative delirium with machine learning in abdominal surgery with comorbidity indices and laboratory values · 2026 · DOI

• In addition, the multiple utilization of small or population-based datasets undermines the generalizability of the models, and lifestyle or metabolic factors established as strongly linked with PCOS are lacking in many works.

  An Explainable Ensemble Machine Learning Framework for PCOS Risk Prediction Using Clinical and Hormonal Data · 2026 · DOI

• indian scenario. The comparison unequivocally demonstrates that the DiabChatBot outperforms DiaBot and Health Care Chatbot in all ML techniques with an accuracy of 98%. in an In 2024 presented a diabetes diagnosis with a CNN and high-frequency ultrasound (HFU). The CNN was applied to the signal's frequency spectra and spectrograms in order to find correlations between signal characteristics and changes in RBC characteristics brought on by glucose levels. Classification accuracy of 0.98 was achieved, confirming the efficacy of the CNN-based approach. In 2025 introduced an integrated intelligent system that combines Optical Character Recognition (OCR) and an LSTM-based chatbot to assist in the management and monitoring of diabetes. The proposed system processes diabetic test reports using OCR to extract key health indicators like HbA1c and VLDL levels. The system accurately classified users and provided personalized guidance using OCR and an LSTM-based chatbot. WSEAS TRANSACTIONS on SIGNAL PROCESSING DOI: 10.37394/232014.2026.22.14Ananthi A., Kirubagari BE-ISSN: 2224-3488161Volume 22, 2026 In the literature review, current techniques have a number of drawbacks, including limited adaptability to diverse populations, lifestyle variations, and environmental factors, particularly outside the Indian context. The system only focuses on diabetes prediction continuous physiological monitoring or personalized feedback, which real-time its effectiveness healthcare and early intervention. To overcome this problem, the DIA-CHAT approach is proposed for diabetes classification.

  DIA-CHAT: Diabetes Classification and Diet Guidance in Pregnant Women using an Intelligent Chatbot · 2026 · DOI

• Future work should focus on validating this framework across larger and more diverse populations, integrating longitudinal data for progression prediction, and exploring advanced ensemble and deep learning methods to further refine diagnostic performance. Our results confirm the strength of the model and the ability to work with mixed data sets.

  Machine Learning Algorithms for Predicting CKD Progression: A Real-World Hospital Dataset Analysis · 2026 · DOI

• Future research should validate the framework using larger hospital datasets, compare oversampling with SMOTE and undersampling methods, perform calibration analysis, test fairness across demographic groups, and incorporate explainable artificial intelligence tools such as SHAP or LIME. A further mathematical extension may combine supervised learning with dynamic disease models, including fractional-order biomedical systems similar to those studied by Okeke et al. [3]. Another promising direction is to formalize robust learning bounds under structured medical noise, extending the precision-dispersion framework of Ozioma et al. [4].

  A Mathematical Model of Analytical Supervised Learning Algorithms for Stroke Prediction Using PySpark: Precision, Dispersion and Random Noise Fluctuation Analysis · 2026 · DOI

• A major strength of this study lies in its use of a large, nationally representative data- set (BDHS 2022) and its application of multiple state-of-the-art machine learning (ML) algorithms, enabling a robust comparative evaluation of predictive performance. The inclusion of SHAP interpretability represents an additional methodological advantage, as it bridges the gap between computational complexity and policy relevance by quan- tifying the contribution of each predictor to model outcomes. This study thus extends prior ML-based public health work by offering an interpretable and generalizable model of women’s mental health care-seeking behavior in a low- and middle-income context. Several limitations must be noted. First, the models’ predictive ability was moderate because DHS surveys did not include psychosocial and attitudinal factors such as per- ceived stigma, social support, and mental health service awareness, which are important Kanchon et al. Discover Public Health (2026) 23:1039 Page 15 of 17 determinants of mental health behavior. Second, the cross-sectional design limits causal inference between predictors and outcomes. Third, the Synthetic Minority Oversam- pling Technique (SMOTE) addressed class imbalance but may introduce synthetic pat- terns that could affect model generalizability. SHAP enhances interpretability but does not imply causality. Additionally, the outcome variable captures whether individuals sought care but does not reflect the timing or frequency of care-seeking, which may limit interpretation regarding early intervention. Finally, since the models were trained and validated on a single-country dataset, their external validity for other populations should be interpreted with caution and may require further validation in different settings. Despite these limitations, the integration of explainable ML into national health research represents a methodological advancement. The study provides a replicable framework for predicting and interpreting health-seeking behaviors using DHS data, contributing to the broader goal of data-driven, equitable, and transparent public health policymaking in low-resource settings.

  Interpretable machine learning for predicting early mental health care-seeking among reproductive-age women in Bangladesh using BDHS 2022 data · 2026 · DOI

• Future work will focus on multi-center prospective external validation, multi-modal data fusion incorporating imaging and genetic susceptibility data, systematic evaluation of class-imbalance mitigation strategies for deep learning architectures, and head-to- head benchmarking against existing RA-specific machine learning frameworks to consolidate the clinical translational value of the proposed pipeline. Constructing a harmonized benchmark dataset for cross-study comparison remains an open methodological challenge that we will pursue in follow-up work. In the present cohort, baseline blood sampling was performed at the patient’s first rheumatology consultation, at which point the rheumatologist had not yet established a definitive RA diagnosis; over the 12-month follow-up, some patients were subsequently classified as RA according to the 2010 ACR/EULAR criteria (39), while others received alternative diagnoses such as osteoarthritis, fibromyalgia, or undifferentiated arthritis.

  Predicting anti-CCP positivity and early rheumatoid arthritis onset from routine laboratory parameters: a SHAP-explained machine learning pipeline · 2026 · DOI

• While this design supports strong internal validity for comparing ML models against the sPESI baseline by minimizing institutional variability in clinical workflows and imaging protocols, the performance of these models in different hospital settings remains to be estab- lished.

  Machine learning-extended sPESI for 1-year mortality prediction in pulmonary embolism · 2026 · DOI

• real-world clinical deployment. These underscore the need for a unified framework that can balance non-linear discriminative features, and ensure generalization without compromising interpretability or scalability.

  Cardiovascular Disease Prediction via Hybrid SVM–SMOTE and Sparse Autoencoder Feature Reduction with Deep MLP Classification · 2026 · DOI

• This paper presented a full-stack ML system for CKD detection and KDIGO stage classification, achieving 95.5% accuracy with 42 ms inference latency. The system's importance, SMOTE- feature strengths—interpretable balanced training, PDF report parsing, offline PWA deployment, and longitudinal patient tracking—make it uniquely suited for Indian public healthcare, particularly in resource-limited settings across Telangana. Future Work: • Phase 1 (Q2 2026): HL7 FHIR API integration for direct EHR connectivity; XGBoost ensemble targeting 97.8% accuracy with SHAP explanations. • Phase 2 (Q3 2026): CNN-LSTM architecture for 12- month biomarker time-series progression risk prediction; 5.2 Validation Test Cases TC-01 (High-Risk Diabetic): 62-year-old male, creatinine = 3.2 mg/dL, albumin low. Output: 94% CKD probability, Stage 4. Top SHAP drivers: creatinine (+2.1), albumin (−1.4). Recommendation: STAT nephrology consult. TC-02 (Medium-Risk Hypertensive): 40-year-old female, BP = 160/90, eGFR = 75. Output: 68% risk, Stage 2 (yellow zone). Post ACE-inhibitor treatment retest: dropped to 22% risk, demonstrating intervention efficacy tracking. TC-03 (False Positive Robustness): 28-year-old, all biomarkers normal except pus cells present. Ensemble voting: RF 18%, XGB 24% — false positive correctly rejected. Demonstrates benefit of soft-voting ensemble over single-model approaches. TC-04 (Multi-Report Trend): Three consecutive reports, Stages 2 → 3 → 3. System outputs: "Worsening" (red badge), "Stable" — enabling longitudinal monitoring. © 2026 The Author(s). Published by IJCOPE Journal. Website: https://ijcope.org/ 4 International Journal of Creative and Open Research in Engineering and Management ISSN: 3108-1754 (Online) Volume 02 Issue 04 April-2026 | Impact Factor: 3.5 federated learning for multi-hospital model improvement without PHI sharing. • Phase 3 (Q4 2026): Flutter cross-platform mobile app with CameraX OCR for in-clinic lab scanning; Apple Health/Google Fit integration for continuous risk scoring. • Phase 4 (Q1 2027): FDA SaMD Class II submission with 21 CFR Part 11 audit trails; bias mitigation framework stratified by age, gender, and ethnicity. • Phase 5 onwards: Population health analytics, SMS- based screening for feature phones, blockchain audit trail for medicolegal validation.

  Detection and Stage Prediction of Kidney Disease using Machine Learning · 2026 · DOI

• The model, in its current form, is more adept at identifying concurrent, undiagnosed dysglycemia than predicting future disease onset, and its added value over simpler FBG-based approaches remains to be established through direct comparative studies.

  An XGBoost-based model for detecting undiagnosed type 2 diabetes using routine physical and lifestyle data from a multi-center Chinese population · 2026 · DOI

• intelligent healthcare prediction, personalized recommendation generation, scalability, data security, and efficient healthcare analytics was used. Thus, the suggested methodology is aimed at the implementation of machine learning algorithms, healthcare techniques, intelligent recommendations, and full-stack web solutions into a unified healthcare decision-support platform. The proposed solution is aimed at the development of preventive healthcare management preprocessing data by analyzing patient-specific physiological and behavioral data in an automated manner without significant involvement of humans in the process. secure place, where Firstly, the registration and authentication process takes authentication procedures based on JSON Web Tokens (JWT) were chosen. Because of the high level of personal information sensitivity, the issue of secure clientserver communication becomes very crucial for healthcare systems. JSON Web Tokens allow for easy and scalable authentication of users, as well as role-based access control and secured client-server communication. token-based authentication approaches are considered one of the best options for implementing safe interaction between a healthcare client and a healthcare server,. Nowadays, such as null value Post-acquisition of the data, the preprocessing related module will undertake operations to transformation. healthcare data cleansing and inconsistent Typically, healthcare data contains values, missing values, duplicates, and noisy inputs from users. Consequently, various preprocessing operations replacement, normalization, label encoding, feature scaling, and outlier elimination will be conducted before the process of training and predicting using the model. Feature engineering operations are conducted to create additional meaningful features that indicate the health status of users in healthcare datasets such as body mass index (BMI), lifestyle score, hydration score, and sleep quality scores,. After completing the process of data preprocessing, the next step involves moving the healthcare dataset to the machine learning prediction layer. Various machine learning algorithms are used for the purpose of prediction and comparison to identify the most suitable algorithm for healthcare risk predictions. They include Logistic regression, Decision tree, Random forest, and XGBoost. Logistic regression algorithm is employed as a baseline classifier because of its interpretability and low computation cost. On the other hand, Ensemble methods such as random forest and XGBoost are relationships and used to discover nonlinear 5 Somit Kumar Yadav, International Journal of Science, Engineering and Technology, 2026, 14:2 complicated variables. interactions between healthcare The generated processed data, prediction information, and recommendation history is saved in the database management system called MySQL. MySQL is used due to its high efficiency in terms of reliability, structured data management, query optimization, and scalability, which make it suitable for healthcare applications. MySQL can store user profiles, health data, prediction results, and recommendation history while ensuring consistency within the system operation. Last but not least, the frontend application serves as an interface for users to visualize their personal health status along with predictions made by the system, as well as get feedback on the generated recommendations. The frontend application is used to make requests to the backend API built on the Flask framework. Asynchronous request handling methods are implemented to ensure smooth interaction between frontend and Flask. Flask framework is used due to its efficient performance, ease of integration with machine learning tools, and flexibility for developing RESTful healthcare APIs. a to Thus, a full stack implementation solution has been facilitate efficient communication created engine, between recommendation generation module, database layer, and frontend application while being able to provide scalable and deployable web-based solutions for healthcare environments.

  AI – Powered Personalization Care Recommendation Engine · 2026 · DOI

• Personalized recommendation engines often fail to incorporate the aspects of adaptability, recommendation ranking based on context, a scalable architecture for deployment, and prevention in healthcare. The vast majority of research done till now is mostly about predictions and not complete healthcare decision support ecosystems. Hence, this research attempts to fill these gaps through AI-powered personalized care recommendation engine. V.

  AI – Powered Personalization Care Recommendation Engine · 2026 · DOI

• Although there are significant advances made in AI healthcare research studies, many researches still concentrate more on disease prediction models and very system implementations. Current research works mostly focus on healthcare condition predictions, symptom categorizations, and diagnoses without paying attention to recommendations, behavioral analysis, and preventive healthcare management. While predictive models bring important insight about health conditions, current research work does not pay much attention to the problems in the practical deployment of such models due to problems in scalability, explainability, real-time deployment, and healthcare data privacy. 4 Somit Kumar Yadav, International Journal of Science, Engineering and Technology, 2026, 14:2 framework There are also many research papers emphasizing the importance of healthcare data security and scalable deployment for an AI healthcare application. Unfortunately, many proposed models in the healthcare industry remain confined only within the theoretical framework and do not provide full-stack implementations which include scalable web application architecture. As a result, the absence of integration of machine learning intelligence and modern-day scalable web the practical architecture greatly healthcare process implementation recommendation applications. Researchers have also indicated that the integration of intelligent healthcare analytics with web technology makes better healthcare adaptable and flexible. development application hampers of engines.

  AI – Powered Personalization Care Recommendation Engine · 2026 · DOI

• healthcare systems outperform their rule-based counterparts recommendation in capabilities. terms of prediction and indicate The use of machine learning technologies including Logistic Regression, Decision Trees, Random Forest, Support Vector Machine and XGBoost technologies has been widely explored in healthcare prediction literature. Machine learning allows computers to analyze heterogeneous healthcare information combining physiological measures, behavior and medical history. The ensemble learning techniques like Random Forest and XGBoost have shown robust results in dealing with nonlinear relations in healthcare and mitigating overfitting problems common to the healthcare data. The past studies have proven that using machine learning can improve preventive analysis of the healthcare by predicting early risk factors connected to the chronic diseases and unhealthy lifestyles. The field of NLP (Natural Language Processing) and conversational healthcare systems have become prominent areas of investigation for intelligent healthcare. The use of conversational AI allows users to communicate with healthcare systems using normal natural language rather than using complicated medical forms only. The study results have indicated that healthcare systems using NLP technology provide better accessibility and usability in comparison with the static interfaces, especially among nontechnical and elder patients. AIdriven conversational healthcare systems have been employed in various healthcare applications including virtual assistants, health monitoring apps and automated patient support systems. on few recommendation Although there are significant advances made in AI healthcare research studies, many researches still concentrate more on disease prediction models and very system implementations. Current research works mostly focus on healthcare condition predictions, symptom categorizations, and diagnoses without paying attention to recommendations, behavioral analysis, and preventive healthcare management. While predictive models bring important insight about health conditions, current research work does not pay much attention to the problems in the practical deployment of such models due to problems in scalability, explainability, real-time deployment, and healthcare data privacy. 4 Somit Kumar Yadav, International Journal of Science, Engineering and Technology, 2026, 14:2 framework There are also many research papers emphasizing the importance of healthcare data security and scalable deployment for an AI healthcare application. Unfortunately, many proposed models in the healthcare industry remain confined only within the theoretical framework and do not provide full-stack implementations which include scalable web application architecture. As a result, the absence of integration of machine learning intelligence and modern-day scalable web the practical architecture greatly healthcare process implementation recommendation applications. Researchers have also indicated that the integration of intelligent healthcare analytics with web technology makes better healthcare adaptable and flexible. development application hampers of engines.

  AI – Powered Personalization Care Recommendation Engine · 2026 · DOI

Most-cited papers in Artificial Intelligence in Healthcare

• Machine Learning-Based Predictive Models for Detection of Cardiovascular Diseases · Diagnostics · 2024 · 252 citations
• Feature reduction for hepatocellular carcinoma prediction using machine learning algorithms · Journal Of Big Data · 2024 · 206 citations
• A literature review of machine learning algorithms for crash injury severity prediction · Journal of Safety Research · 2021 · 199 citations
• Machine learning and deep learning-based approach in smart healthcare: Recent advances, applications, challenges and opportunities · AIMS Public Health · 2024 · 183 citations
• Review of multimodal machine learning approaches in healthcare · Information Fusion · 2024 · 167 citations
• Classification models combined with Boruta feature selection for heart disease prediction · Informatics in Medicine Unlocked · 2024 · 119 citations
• A proposed technique for predicting heart disease using machine learning algorithms and an explainable AI method · Scientific Reports · 2024 · 116 citations
• Implementing machine learning in medicine · Canadian Medical Association Journal · 2021 · 115 citations
• Enhancing heart disease prediction using a self-attention-based transformer model · Scientific Reports · 2024 · 112 citations
• Cardiovascular disease diagnosis: a holistic approach using the integration of machine learning and deep learning models · European journal of medical research · 2024 · 108 citations

Most recent work

• Forecasting Students' Academic Performance in Educational Data Using Machine Learning Techniques · International Journal of Information and Communication Technology Education · 2026
• Advanced Healthcare Analytics Using AI, ML, and IoT: A CNNBased Algorithmic Approach · International Journal of Drug Delivery Technology · 2026
• Cardiovascular Disease Prediction via Hybrid SVM–SMOTE and Sparse Autoencoder Feature Reduction with Deep MLP Classification · Informatica · 2026
• Work-Life Conditions as the Primary Determinant of Seafarer Mental Health: An Explainable Machine Learning Analysis · INQUIRY The Journal of Health Care Organization Provision and Financing · 2026
• A Machine Learning-Based System for Early Diabetes Prediction using Medical Data · International Scientific Journal of Engineering and Management · 2026
• AI-Driven Treatment Response Prediction for Chronic Disease Management Using Generative Artificial Intelligence and Large Language Models · International Journal Of Recent Trends In Multidisciplinary Research · 2026
• Mathematical Optimization and Ensemble Learning for Hypertension Risk Prediction: A Feature-Driven Stacked Framework with Interpretability · Contemporary Mathematics · 2026
• Early Type 2 diabetes risk prediction using explainable machine learning in a two-stage approach · Frontiers in Digital Health · 2026
• Heart Risk Analysis via Health Factors · International Journal of Mathematics And Computer Research · 2026
• An Ensemble Machine Learning Approach for Early Liver Disease Prediction · International Scientific Journal of Engineering & Management · 2026

Related topics in Health Professions