Hospitals hesitate to pool perioperative records for AKI models because of privacy rules and heterogeneous EHR schemas. We train a federated gradient-boosted ensemble across seven cardiac centers that never exchange raw features, synchronizing only encrypted parameter updates. External validation on a held-out site yields AUROC comparable to a centrally trained baseline while reducing false-positive alerts that drive unnecessary nephrology consults during the first 48 postoperative hours.

Clinic-based physiotherapy schedules conflict with return-to-work timelines after ACL reconstruction. We compare a video-supervised home program with usual in-person care on twelve-week Lysholm scores and session adherence. Tele-rehabilitation matches functional gains when weekly live check-ins are retained, while fully asynchronous apps show lower adherence after week six. Patient interviews emphasize that real-time form correction, not merely exercise videos, sustains engagement during demanding occupational schedules.

Discharge summaries often leave unresolved drug discrepancies that surface only after patients return home. We evaluate a structured pharmacist reconciliation visit within seventy-two hours of discharge for adults with polypharmacy. The intervention arm shows fewer clinically significant discrepancies and a modest reduction in thirty-day unplanned readmissions, concentrated among patients with heart failure or anticoagulants. Process metrics indicate that physician callback pathways determine whether identified issues are corrected before harm occurs.

Specialized imaging biomarkers are rarely available in general diabetic foot clinics. We train gradient-boosted models on routinely recorded ulcer size, infection grade, glycemic control, and vascular scores to predict twelve-week non-healing. External validation at a second clinic maintains discrimination above chance and identifies infection grade plus ankle-brachial index as dominant contributors. Calibration plots improve after isotonic adjustment, supporting use as a triage aid rather than a standalone surgical decision tool.

Existing moral injury scales are lengthy for repeated staff wellness screening during prolonged surge cycles. We adapt and validate an eight-item inventory in frontline nurses recruited across medical and intensive care units. Exploratory and confirmatory analyses support a two-factor structure separating betrayal-related and self-transgression items, with acceptable internal consistency and expected correlations with burnout and intention to leave.

Subsidy programs aim to cut indoor smoke exposure, yet adoption and sustained use vary widely across neighborhoods. We link clinic records of lower respiratory infections in children under five with household fuel surveys before and after a staged clean cookstove subsidy. Communities with higher sustained LPG use show larger declines in clinic visits, while mixed-fuel stacking attenuates benefits. Qualitative interviews highlight refill cost and supply reliability as stronger barriers than stove preference.

Pooling labeled radiographs across hospitals often stalls on privacy and governance barriers. We train a shared triage network with federated averaging across four sites that never exchange raw images, evaluating urgent-finding recall against locally trained baselines. Federated models approach centralized performance for pneumothorax and consolidation flags while narrowing site-specific overfitting, although non-identical acquisition protocols still require periodic local fine-tuning.