Medication recommendation systems generate either fixed top-ω medication lists or threshold-based multilabel predictions. However, these approaches do not control for the risk of omitting medications from true prescription set. This study proposes Coverage-Aware Medication Recommendation (CAMR), a framework that applies Conformal Risk Control (CRC) to a multilabel classifier to generate patient-specific set-valued medication recommendations. CAMR calibrates thresholds for a user-defined False Negative Rate (FNR)-based risk target. The framework was evaluated using a seven-year retrospective outpatient cohort from a tertiary hospital. The baseline classifier achieved an empirical coverage of 0.7962, an F1-score of 0.8276, and an average set size of 2.65 medications. At target risk levels of 0.20, 0.15, and 0.10, corresponding to nominal coverage levels of 80%, 85%, and 90%, CAMR achieved pooled empirical coverage rates of 0.8029 (95% CI 0.8010–0.8048), 0.8493 (0.8476–0.8509), and 0.8961 (0.8947–0.8973), with F1-scores of 0.8275, 0.8063, and 0.7137 and average set sizes of 2.70, 3.17, and 4.33 medications. A Top-5 strategy achieved similar coverage (0.8019) only with five medications and a lower F1-score of 0.5844. Conventional conformal prediction baselines required larger sets to achieve comparable coverage. Performance degradation was concentrated in infrequent medication classes and visits involving larger medication combinations. Sensitivity analyses showed that class-weighted and group-conditional calibration corrected under-coverage in rare classes and in departments with long prescriptions, while coverage remained close to the target over the test period. Overall, pooled empirical coverage approached nominal targets. The theoretical guarantee remains conditional on exchangeability, which may be affected by repeated visits and temporal distribution shift.
Timely and efficient access to guideline-concordant care is a decisive determinant of outcomes and costs in chronic immune-mediated diseases such as Rheumatoid Arthritis (RA). Chronic-care delivery in Colombia is fragmented, with barriers to access, diagnosis, treatment, and follow-up that generate productivity loss and disability outcomes while inflating costs. This study aims to apply process mining technologies to discover, model, and evaluate RA care pathways in Colombia’s health system. We used a nationwide administrative claims and provider transactions to discover the actual care pathway, to quantify deviations from guideline-concordant should-be pathways, and to identify bottlenecks, delays, and coverage gaps. We used a suite of complementary flow, performance, conformance, and variant analysis to expose multiple discontinuities that routine aggregates would overlook. These analysis allowed us to quantify timelines and service bundles in the reconstructed core treat-to-target pathway for patients with RA. The longest interval identified between procedures was 9.89 months and may indicate fragmented access to supportive or interdisciplinary services. Nutritional support (0.6%) and serum IgM rheumatoid-factor testing (0.2%) were infrequently recorded relative to the national RA risk-management indicators used as reference standards. These findings demonstrate the potential of process mining to identify documented pathway patterns and potential care gaps that may inform further clinical and organizational evaluation and pathway redesign.
Clinical alert systems are commonly evaluated using predictive performance and decision-analytic measures that treat all true-positive alerts within a prediction horizon as beneficial. For time-sensitive decisions, this assumption may be too permissive because a correct alert may arrive too late to support action. This study introduces Actionable Net Benefit (ANB), a timing-aware extension of net benefit that counts a true-positive alert as beneficial only when it satisfies a prespecified minimum lead-time requirement. ANB was evaluated using a retrospective inpatient falls case study comprising 58,803 discharge-aware patient-day alert opportunities from 5451 encounters, including 96 positive 24-hour windows. A ridge-penalized logistic model was selected through encounter-stratified cross-validation and evaluated using nested encounter-level out-of-fold predictions. The model achieved an out-of-fold AUROC of 0.810 and AUPRC of 0.00715, compared with 0.560 and 0.00398 for the Hester-Davis comparator. Operational capacity was applied within each 7:00 AM alert cycle and was separated from the threshold probability used to weight false-positive burden. Under the reference analytic scenario of 5% per-cycle capacity, the simulated policy generated 2336 alerts and selected 19 positive windows, of which 13 remained actionable under the reference 6-hour lead-time requirement. At pt=0.005, ANB decreased from 1.502 per 10,000 alert opportunities at L=0 to 0.277 at L=6. The corresponding encounter-clustered bootstrap 95% intervals were 0.081 to 3.850 and −0.979 to 2.084, and the paired change was −1.225 (−2.592 to −0.409). At L=6, ANB ranged from 1.469 to −2.124 across the lower and higher threshold-probability scenarios. Schedule-specific recalibration improved probability alignment for 12-hour and 6-hour horizons, while the static daily predictor design produced lower ANB under more frequent schedules. ANB provides a structured approach for evaluating whether candidate alert policies remain useful after operational capacity, preference weighting, and response-time requirements are considered.
Automated diagnosis of retinal diseases from fundus images has become an important application of deep learning (DL) for supporting timely clinical diagnosis and treatment. Although existing DL models have achieved high predictive accuracy, many offer limited interpretability, lack rigorous validation of their explanations, and do not integrate ensemble learning with explainable artificial intelligence (XAI) approaches, thereby reducing their clinical utility. This paper proposes XDR-Net, an explainable deep learning framework for multi-disease retinal diagnosis that integrates accurate prediction with disease-specific and quantitatively evaluated explanations. The framework makes three main contributions. First, it combines four complementary convolutional neural networks (CNNs), VGG19, DenseNet121, MobileNetV2, and AlexNet, into a heterogeneous ensemble to improve classification performance on the evaluated dataset. Second, it introduces a disease-aware explanation strategy that selects the most appropriate explanation method according to disease characteristics, using Gradient Weighted Class Activation Mapping Plus (Grad-CAM++) for structural abnormalities and patch-based Multiple Instance Learning (MIL) for lesion-centered diseases, thereby producing disease-specific model explanations. Third, it incorporates a quantitative validation module that evaluates explanation reliability using the confidence drop and energy concentration metrics to quantitatively assess model prediction sensitivity to highlighted image regions. Experiments on fundus image dataset achieved a classification accuracy of 94.02% across diabetic retinopathy, glaucoma, cataract, and normal cases. Comparative evaluations with the evaluated baseline approaches showed competitive predictive performance and quantitative analysis of the generated explanations. These findings indicate that XDR-Net provides competitive classification performance and disease-specific model explanations on the evaluated dataset.
This study presents a data-driven framework for operating room (OR) planning and scheduling that integrates machine learning-based surgery duration prediction with an integer linear programming (ILP) scheduling model under realistic hospital constraints. Efficient OR management remains a major operational challenge due to the stochastic nature of surgical durations and the complexity of coordinating hospital resources. The proposed approach is based on a rolling-horizon planning methodology for elective surgeries. Surgery durations are predicted using machine learning models trained on a large real-world dataset containing more than 90,000 surgeries across 540 procedures and 744 surgeons. The predicted durations are then incorporated into an ILP model that considers surgeon availability, equipment limitations, and OR-specific operational constraints. The framework is evaluated against the hospital’s current duration-estimation approach as well as a benchmark schedule based on actual surgery durations. Results show substantial operational improvements, including a reduction of more than 60% in unused OR capacity with only a modest increase in overtime. In addition, the gap from the benchmark schedule is reduced by 71% compared with the hospital’s current estimation method. The results further demonstrate an important operational trade-off: a modest increase in overtime can lead to a substantial reduction in unused OR capacity. This trade-off provides hospital managers with practical flexibility in balancing OR utilization and operational risk according to institutional priorities. Overall, the study highlights the value of integrating predictive analytics and optimization methods to support data-driven OR planning in large-scale hospital environments.
