Wearable based monitoring and self-supervised contrastive learning detect clinical complications during treatment of Hematologic malignancies

**Background:** Treatment of patients with hematological malignancies is associated with a high incidence of clinical complications such as infections, cardiac events, and immunologic dysregulations. Early recognition and therapeutic intervention are critical, as delayed intervention increases morbidity and mortality. With increasing oncological treatments shifting to outpatient settings, complication detection relies primarily on patient self-assessment, which often leads to delayed recognition. Remote patient monitoring (RPM) with medical wearables offers non-invasive, continuous real-time monitoring of vital signs and physical activity. This study aimed to evaluate whether wearable-based RPM combined with a self-supervised contrastive deep learning model could detect and predict serious clinical complications (SCC) in patients undergoing treatment for hematological malignancies.

**Methods:** This open-label, single-arm, single-center, investigator-initiated cohort study was conducted at the Department of Hematology, Oncology, and Clinical Immunology of the University Hospital Düsseldorf, Germany. A convenience sample of 79 patients was recruited: 54 inpatients (IC) and 25 outpatients (OC). Inclusion criteria were age ≥18 years and treatment with expected hematotoxicity according to CTCAE grade 4. Patients wore a commercially available CE-marked wearable (Everion, Biovotion AG, Switzerland) that continuously monitored heart rate, temperature, respiratory rate, and physical activity. Two wearables were assigned per patient for alternate use. Non-hematological SCC were defined as CTCAE (v4.03) grade ≥3. Clinical documentation was independently reviewed by two investigators. Hours with documented SCC plus a 48-hour buffer before and 24-hour buffer after were annotated as non-regular; remaining hours were regular. A ResNet architecture with 24 residual blocks was trained using self-supervised contrastive learning. The SCC-Score was defined as one minus the maximum cosine similarity between a test hour and the reference set. Ten-fold cross-validation was performed. Primary outcomes were detection and prediction of clinically documented SCC.

**Key Results:** A total of 140 SCC events were documented. Cumulative incidence of SCC was 90.7% in IC and 48.0% in OC. Infectious SCC accounted for 65.0% of all SCC. Wearable data were recorded for 24,047 hours for IC (median 457.4 hours per patient) and 7,187 hours for OC (median 315.5 hours per patient). In the patient-non-specific approach, the SCC-Score for regular vs. non-regular hours was significantly different (total cohort: 0.096 ± 0.038 vs. 0.146 ± 0.051, p<0.0001), with AUROC of 0.80 ± 0.01 for all SCC and 0.82 ± 0.01 for infectious SCC. The patient-specific approach showed substantially improved performance: AUROC of 0.93 ± 0.01 for all SCC and 0.94 ± 0.01 for infectious SCC. At ~95% sensitivity, specificity was 61.6% for all SCC and 66.7% for infectious SCC in the patient-specific approach. Per-day SCC-Score further increased AUROC by ~10% (IC 0.85, OC 0.89). For infectious SCC, the SCC-Score showed a transient increase beginning approximately 48 hours before clinical diagnosis, with a steeper slope before diagnosis than the decline post-diagnosis. The averaged z-score for the most common SCC ('Infections and infestations—other') was 2.21 (n=66).

**Clinical Implications:** This study provides proof of principle that wearable-based RPM combined with a self-supervised deep learning model can detect and predict SCC up to 48 hours before clinical diagnosis in patients with hematological malignancies. The patient-specific approach significantly outperformed the patient-non-specific approach, demonstrating the importance of individualized reference sets. The ability to predict infectious SCC early could enable timely therapeutic interventions, potentially reducing morbidity and mortality. The SCC-Score provides a single actionable value that can be translated into clinical decisions, with adjustable thresholds based on clinical context. Future applications could include real-time analysis, demand-driven visits reducing clinician workload, and decreased frequency of blood sampling. Limitations include the small sample size (though largest trial in this population), potential under-detection of lower-grade complications, and variability in how different SCC types affect vital signs. Prospective confirmatory studies are needed to document clinical benefit in practice.