Lee, Seungwoo

Abstract Introduction Distinct temperature patterns have long been recognized to correlate with fevers of differing etiologies. While the use of wearable sensors for high-frequency temperature monitoring (HFTM) on a near minute-by-minute basis has been shown to detect fevers earlier than standard-of-care nursing vital sign assessments in hospitalized patients, leveraging these high-resolution datasets to computationally identify unique digital signatures for real-time diagnosis of underlying fever etiology has not been widely explored. Diagnostic uncertainty is common in patients undergoing hematopoietic stem cell transplantation (HCT), with only 20–30% of febrile neutropenic episodes being microbiologically documented. We hypothesized that unique temperature patterns extracted from HFTM data collected during episodes of febrile neutropenia could be used to develop a supervised machine learning classifier capable of accurately predicting underlying fever etiology in HCT patients. Methods We analyzed 68 clinically independent fever episodes recorded in HCT patients (n=90) outfitted with an FDA-cleared wireless temperature sensor (TempTraq®, BlueSpark Technologies) that measured axillary temperature every 2 minutes throughout hospitalization. Time-series features were extracted from temperature traces spanning 1 hour before to 3 hours after fever onset and used to train a suite of machine-learning models to distinguish engraftment fevers from other fever etiologies. Model training and evaluation were performed using repeated stratified 5-fold patient-level cross-validation, yielding 100 train-test evaluations. Results Among all classification models, the logistic regression classifier provided the best overall performance and interpretability, achieving 94% specificity (95% CI, 0.84–1.0) for identifying engraftment fevers with a mean AUROC of 0.88 ± 0.10. Feature importance analysis demonstrated that both clinical variables and HFTM-derived temperature dynamics contributed to model performance, with a strong reliance on time-series features captured within the first 4 hours of fever onset. Conclusion Our study provides a demonstration that continuous temperature data collected from patients outfitted with wearable sensors can be leveraged not only for early fever detection but also for machine learning–based diagnosis of fever etiology. These findings suggest that dynamic temperature patterns contain clinically meaningful physiologic information that with further studies could support real-time diagnostic decision-making and guide safe de-escalation of empiric antibiotics during febrile neutropenia in patients undergoing intensive cancer therapy.

592 Background: While not all fevers occurring during hematopoietic stem cell transplantation (HCT) are dangerous, episodes of febrile neutropenia tend to be managed aggressively given the significant mortality and morbidity associated with infectious complications of transplantation. This can potentially lead to unnecessary diagnostic testing, antibiotic treatment, and prolonged hospital stays for non-infectious fever etiologies such as engraftment. Our group has previously demonstrated that high-frequency body temperature monitoring (HFTM) using a wearable device detects fevers earlier in transplant recipients (Flora et al, Cancer Cell, 2021). We hypothesized that applying supervised machine learning to HFTM data collected from inpatient allogeneic HCT patients would allow for accurately distinguishing between non-infectious and infectious etiologies of fever and could potentially guide safe de-escalation of antibiotics. Methods: We analyzed 74 independent fever events detected in HCT patients (n=42) outfitted with an FDA-cleared, wireless sensor measuring axillary temperatures every 2 minutes for the duration of their respective hospitalizations. After data filtering, independent fever events were then retrospectively grouped by fever etiology using physician-determined annotation as previously described (Flora et al., 2021). A split ratio of 80:20 was used to create training and test sets from the parent data set, consisting of 59 and 15 fever events for each split, respectively. Training sets were used to create a classification algorithm identifying bacterial infections requiring antibiotics using the machine learning model XGBoost. The performance of this model was then evaluated using test sets of the remaining fevers (n=15). Results: Averaging over 1000 test sets of 15 independent fevers comprising 5 bacterial infection-related fevers and 10 non-infectious fevers, our model correctly identified 82% (mean recall) of cases requiring antibiotics with a mean precision of 0.41. The model also correctly identified 42% of non-infectious fevers (mean true negative rate, 0.42) that would otherwise have been treated with unnecessary antibiotics. The mean AUCPR was 0.58, with the best performing individual model achieving an AUCPR of 0.97. Conclusions: The preliminary results of this study demonstrate that our classification model shows promise for differentiating fevers requiring antibiotics due to bacterial infections from other non-infectious fever etiologies. Prospective studies further developing supervised machine learning models for subcategorization of febrile neutropenia could pave the way for safely de-escalating antibiotics in HCT and other immunocompromised patients undergoing cancer treatment monitored using HFTM.