Machine learning
Machine learning (ML) is a field of artificial intelligence concerned with computational methods that identify patterns in data and use them to generate predictions, classifications, or decisions.
Machine learning (ML) is a field of artificial intelligence concerned with computational methods that identify patterns in data and use them to generate predictions, classifications, or decisions. In contrast to rule-based systems, ML models estimate relationships from examples, such as clinical variables, laboratory measurements, medical images, text, or molecular profiles. Common approaches include Logistic Regression, support vector machine, random forest, XGBoost, and LightGBM; deep learning represents a related family of models that uses multilayer neural networks. Model performance may be assessed with measures such as Accuracy, sensitivity, specificity, F1 score, and the Area Under the Receiver Operating Characteristic Curve.
In biomedicine, ML is used for risk stratification, diagnosis, prognosis, image segmentation, clinical decision support, and analysis of high-dimensional biological data. Its potential value lies in integrating multiple predictors—including static patient characteristics and changing biomarkers—to estimate individualized risks. Interpretability methods such as SHapley Additive exPlanations and XGBoost-SHAP can help describe how input variables contribute to a prediction, while Decision Curve Analysis can assess the potential clinical usefulness of a model. ML is therefore a computational tool rather than a biological treatment or pharmacological mechanism; its clinical significance depends on data quality, validation, interpretability, and successful integration into care.
Rebuilt from PubMed 18 Sept 2026 · no new papers today
Where the papers sit
15 papers study machine learning directly. Those 15 are one subject: Clinical Risk Prediction. Clinical risk prediction is expanding from fixed prognostic scores toward interpretable, flexible-horizon models that combine clinical, radiomic and longitudinal data. Applications span dementia, cancer, postoperative complications, thrombosis, ICU weakness and treatment selection, with multicenter validation and decision support recurring. No way of splitting those 15 scores better than chance. 1 paradigm shift and 1 new direction follow.
Clinical machine-learning risk prediction need not be static or tied to a single fixed horizon
Dementia early detection and onset prediction and hematological-malignancy-associated peripherally inserted central catheter thrombosis are treated in separate studies as problems inadequately served by fixed-window or static risk models. The dementia study instead combines diagnosis with prediction across horizons from 24 to 120 months, while the thrombosis study combines baseline clinical data with dynamic biomarkers for individualized risk assessment (PMIDs 42742759, 42618846). Together, they replace the assumption that clinical ML produces a one-time, fixed-horizon score with a model of prediction that can represent changing time horizons and patient states.
Machine learning can be used to select sequential treatment policies rather than only predict clinical outcomes
Dynamic treatment regimes with ordinal outcomes are the subject of a Bayesian machine-learning study that introduces a method for estimating optimal sequences of decision rules from patients’ treatment histories and evolving disease status, while quantifying uncertainty 42639730Aug. This gives ML a role not represented by the other papers’ detection, segmentation, classification, risk-prediction, or implementation analyses: it is used to construct and evaluate adaptive treatment strategies, changing the task from forecasting an outcome to choosing how treatment should unfold.
Recent Findings on Machine learning
Clinical Risk Prediction: Clinical prediction models increasingly use routine clinical variables to stratify risks ranging from dementia and postoperative complications to mortality, thrombosis, recurrence and newborn screening outcomes 42742759Sep42642795Aug42618846Aug42530089Jul42527084Jul42225202Jun. Dynamic biomarkers, prediction horizons and treatment history are extending models toward individualized surveillance and optimal dynamic treatment regimes 42742759Sep42618846Aug42639730Aug. SHapley Additive exPlanations, calibration, decision curve analysis and web-based tools are making predictions more interpretable and accessible, while responsible AI studies examine fairness and ethical oversight 42742759Sep42642795Aug42618846Aug42530089Jul42526023Jul. Performance remains uneven: XGBoost outperformed logistic regression for PICC-related thrombosis, whereas radiomic features added limited value to clinical predictors, and local recurrence prediction showed modest discrimination 42618846Aug42679396Sep42527084Jul. Multicenter external validation, prospective evaluation, recalibration, workflow redesign and clinician engagement now guide development, including planned ensemble modelling for colorectal cancer screening among people with HIV 42716689Sep42150783May42150828May.
Written from 15 PubMed abstracts, each one cited by PMID above. Published: 2026-08-29. Last written: 2026-09-17 by GPT. Drafted by language models from published abstracts; not medical advice.