Development and Internal Validation of a Nurse-Led Sepsis Risk Score Using AI-Enabled Remote Patient Monitoring Data in Adult Febrile Neutropenia: A Nested Case–Control Study

Background/Objectives: Febrile neutropenia (FN) in hematologic malignancies carries a high risk of sepsis. We developed a regression model and a simplified nurse-led sepsis risk score using artificial intelligence (AI)-enabled remote patient monitoring (RPM) measurements and routine clinical and laboratory variables. Methods: A nested case–control study included 120 adult FN episodes (40 Sepsis-3-defined cases within 14 days of FN onset; 80 matched controls) in a hematology/bone marrow transplantation unit (BMT) in India (December 2021–December 2024). Heart rate and respiratory rate were obtained by ballistocardiography; blood pressure (BP), peripheral oxygen saturation (SpO2), and temperature were obtained using connected probes. Only a single set of FN-onset measurements was analyzed; longitudinal RPM trends and platform-generated alerts were excluded. Conditional logistic regression identified independent predictors and supplied coefficients for an integer-weighted score. Regression-model validation used two 1000-resample bootstrap analyses: one holding cut-offs and predictors fixed, the other repeating cut-off determination and backward selection. Integer-score discrimination was assessed separately. Results: Six predictors—tachycardia, hypotension, ≥2 comorbidities, absolute neutrophil count <99.5 cells/μL, chills, and elevated procalcitonin—formed an 8-point score. The regression model area under the receiver operating characteristic curve (AUC) was 0.847, with fixed-specification optimism of 0.028 and a corrected AUC of 0.819. Repeating cut-off determination and predictor selection yielded an apparent AUC of 0.855, optimism of 0.036, and corrected AUC of 0.819. The integer-score apparent AUC was 0.654 (95% CI 0.567–0.740), without optimism correction. Conclusions: Regression-model discrimination remained similar after accounting for cut-off determination and predictor selection. Integer-score performance remains uncorrected for optimism and requires independent temporal validation before clinical use. RPM’s AI functions and continuous-monitoring benefits were not evaluated. Temporal validation is planned in Phase II. Trial registration: Clinical Trials Registry of India (CTRI/2025/06/088290).