Background Venous thromboembolism (VTE) is a chronic disorder with a significant health and economic burden. Several VTE-specific clinical prediction models (CPMs) have been used to assist physicians in decision-making but have several limitations. This systematic review explores if machine learning (ML) can enhance CPMs by analyzing extensive patient data derived from electronic health records. We aimed to explore ML-CPMs' applications in VTE for risk stratification, outcome prediction, diagnosis, and treatment. Methods Three databases were searched: PubMed, Google Scholar, and IEEE electronic library. Inclusion criteria focused on studies using structured data, excluding non-English publications, studies on non-humans, and certain data types such as natural language processing and image processing. Studies involving pregnant women, cancer patients, and children were also excluded. After excluding irrelevant studies, a total of 77 studies were included. Results Most studies report that ML-CPMs outperformed traditional CPMs in terms of receiver operating area under the curve in the four clinical domains that were explored. However, the majority of the studies were retrospective, monocentric, and lacked detailed model architecture description and external validation, which are essential for quality audit. This review identified research gaps and highlighted challenges related to standardized reporting, reproducibility, and model comparison. Conclusion ML-CPMs show promise in improving risk assessment and individualized treatment recommendations in VTE. Apparently, there is an urgent need for standardized reporting and methodology for ML models, external validation, prospective and real-world data studies, as well as interventional studies to evaluate the impact of artificial intelligence in VTE.

Ischemic stroke is a medical emergency that requires hospitalization and occasionally, specialized care at the Intensive Care Unit. Mortality prediction in the ICUs has been a challenge for intensivists, since prompt identification could impact medical clinical practices and allow efficient allocation of health resources in the ICUs, which are extremely restricted, especially in the era of COVID-19 pandemic. Clinical decision support systems based on machine learning algorithms are taking advantage of the vast amount of information available in the ICUs and are becoming popular in the medical predictive analysis. This study aims to explore the feasibility of interpretable machine learning models to predict mortality in critically-ill patients suffering from stroke. To do so, a vast variety of clinical and laboratory information stored in the electronic health record, are pre-processed to allow taking into account the temporal characteristics of a patient’s stay. An 8-hour sliding observation window was utilized. For the experimental evaluation we used the Medical Information Mart for Intensive Care Database (MIMIC-IV). Results indicate sufficient ability to predict mortality at the end of a given day during the patient’s stay. Moreover, attribute evaluation highlights the important indicators to consider when following up with a patient.

Background and Objectives: Irregularly sampled intensive care unit (ICU) electronic health records (EHRs) are difficult to transform into reproducible patient-level representations for outcome prediction and external validation. We propose a deterministic multi-temporal tensorization framework for representing early ICU trajectories in stroke patients and evaluate its utility for ICU length-of-stay (LOS) regression and in-hospital mortality classification under strict no-refit external validation.Methods: Clinical trajectories from the first 48 hours after ICU admission were summarized using expanding observation windows of 1, 2, 3, and 4 hours. Within each window, 77 clinician-selected clinical and laboratory variables were aggregated using mean, median, minimum, and maximum values. Age, gender, and race were encoded as time-invariant demographic variables and replicated across the temporal axis. The resulting summaries were aligned to a common 48-step timeline and stacked into a four-channel tensor representation. Tensor-derived tabular features were used to train XGBoost models for LOS prediction and mortality classification. Model development was performed on MIMIC-IV using patient-level splits, while eICU was used as a fully held-out external cohort without retraining, threshold optimization, or recalibration. Performance and calibration were assessed at the patient level, while SHAP-based interpretability was performed on the internal MIMIC-IV test set.Results: For LOS prediction, internal MAE/RMSE were 1.001/1.440 days with R^2=0.590$, compared with 1.199/1.841 days (R^2=0.227$) externally. For mortality prediction, ROC-AUC was 0.925 internally and 0.901 externally. PR-AUC decreased from 0.830 to 0.414 but remained well above the cohort-specific no-skill baselines of 0.204 and 0.064. At a fixed threshold of 0.5, precision decreased from 0.900 to 0.409, sensitivity was 0.554 and 0.581, and specificity was 0.984 and 0.943. Calibration analysis showed external drift, with overestimation of mortality risk (calibration intercept -1.451; slope 0.619) and underestimation of LOS (mean residual bias +0.489 days). Internal-test SHAP analysis identified clinically plausible predictors, including Glasgow Coma Scale, oxygenation and respiratory support variables, and hematologic and inflammatory markers.Conclusions: Multi-temporal tensorization provides a reproducible computing methodology for transforming irregular ICU EHRs into structured representations suitable for scalable machine learning and cross-database validation. Preserved discrimination did not imply transportable calibration, supporting the need for local calibration assessment before clinical deployment.