Distinguishing Infection- Vs. Non-Infection-Related Febrile Neutropenia in HCT Patients By Machine Learning Analysis of Continuous Body Temperature Data Collected Using a Wearable Sensor

Introduction Febrile neutropenia (FN) occurs commonly in patients undergoing Hematopoietic Cell Transplantation (HCT). We have shown that high-frequency temperature monitoring (HFTM) using a wearable device detects fevers in hospitalized HCT patients earlier (Flora et al, Cancer Cell, 2021). The “continuous data” collected by HFTM (e.g., every 2 minutes) raises the possibility that temperature patterns observed around the time of fever onset could predict whether a FN episode will be associated with documentable infection or not. Such a prediction could expedite the identification of FN patients without infection, potentially enabling sooner antibiotic cessation and earlier hospital discharge. Objective Our objective was to apply unsupervised machine learning to HFTM data from an observational study of allogeneic HCT inpatients who developed fevers, to test the hypothesis that temperature patterns could differentiate fevers due to infection vs. without discernible infection. This involved analyzing data collected +/-4 hours of fever onset time using an FDA-cleared, wireless temperature sensor placed in the axilla. Methods We analyzed 24 fever events from HCT inpatients, comparing HFTM data with physician-determined labeling of events as associated with infection or not. After data filtering, interpolation and smoothing, we selected independent fever events using our published criteria (Flora et al.). We then grouped fevers based on temporal patterns in the HFTM data +/-4 hours around fever start time, by applying k-means clustering with an optimal cluster number of n=4, to a dynamic time-warping distance calculated from the time series data. Results Temperature traces in each cluster had a distinctive pattern of temperature dynamics (Figure 1 (top)). Cluster 1 fevers exclusively correspond to non-infectious causes, cluster 4 exclusively to infection, while cluster 2 favored non-infectious causes and cluster 3 favored infectious causes (Figure 1 (bottom). After labeling different combination of clusters as infection vs. non-infection, we observed an overall AUC of 0.88 on ROC curve analysis for identification of fever associated with infection. Conclusion Our findings suggest that HFTM data's temporal patterns around fever events can differentiate infection and non-infection causes. This provides rationale for further research using bigger, prospective studies and supervised machine learning methods to train classification models, including clinical data from the electronic health record (EHR) as well. It paves the way for future trials using HFTM to identify patients in whom antibiotics can be stopped earlier (e.g., after 24 hours) and/or who could be discharged sooner.