Development and Clinical Validation of a Multimodal AI Framework to Predict Persistent Fever During Antibiotic Therapy in Hospitalized Patients with Cancer

Among hospitalized cancer patients, persistent fever despite early intravenous antibiotic therapy often triggers treatment escalation and additional diagnostics. Earlier identification of patients at low risk for ongoing fever could support more selective clinical decision-making. We developed a feature-level multimodal AI framework to predict whether patients would remain febrile at the 48–72-hour antibiotic reassessment window. Using structured clinical variables, time-series forecasts, note-derived phenotypes, and CT-derived features, we fit the model in a cancer patient cohort from University Hospital Essen and evaluated temporal holdout and external transportability to two pooled regional hospitals and an ICU cohort (MIMIC-IV). Compared with baseline strategies, the multimodal model achieved the highest internal discrimination (AUC 0.83, 95% CI 0.81–0.86) and maintained performance in external ICU validation (MIMIC-IV AUC 0.79, 95% CI 0.76–0.83). In a clinician validation study, the multimodal model outperformed unaided clinicians in both non-AI scenarios (Trial 1 balanced accuracy: 0.77 vs 0.60; Trial 2: 0.79 vs 0.57), while AI assistance improved clinician balanced accuracy to 0.71 (95% CI 0.67–0.75). Overall, this multimodal framework demonstrates potential to support 48-hour antibiotic reassessment and stewardship-oriented decision-making, although further prospective implementation studies are needed before clinical deployment.