Frontiers in Digital Health (Jul 2021)
Identifying Elevated Risk for Future Pain Crises in Sickle-Cell Disease Using Photoplethysmogram Patterns Measured During Sleep: A Machine Learning Approach
Abstract
Transient increases in peripheral vasoconstriction frequently occur in obstructive sleep apnea and periodic leg movement disorder, both of which are common in sickle cell disease (SCD). These events reduce microvascular blood flow and increase the likelihood of triggering painful vaso-occlusive crises (VOC) that are the hallmark of SCD. We recently reported a significant association between the magnitude of vasoconstriction, inferred from the finger photoplethysmogram (PPG) during sleep, and the frequency of future VOC in 212 children with SCD. In this study, we present an improved predictive model of VOC frequency by employing a two-level stacking machine learning (ML) model that incorporates detailed features extracted from the PPG signals in the same database. The first level contains seven different base ML algorithms predicting each subject's pain category based on the input PPG characteristics and other clinical information, while the second level is a meta model which uses the inputs to the first-level model along with the outputs of the base models to produce the final prediction. Model performance in predicting future VOC was significantly higher than in predicting VOC prior to each sleep study (F1-score of 0.43 vs. 0.35, p-value <0.0001), consistent with our hypothesis of a causal relationship between vasoconstriction and future pain incidence, rather than past pain leading to greater propensity for vasoconstriction. The model also performed much better than our previous conventional statistical model (F1 = 0.33), as well as all other algorithms that used only the base-models for predicting VOC without the second tier meta model. The modest F1 score of the present predictive model was due in part to the relatively small database with substantial imbalance (176:36) between low-pain and high-pain subjects, as well as other factors not captured by the sleep data alone. This report represents the first attempt ever to use non-invasive finger PPG measurements during sleep and a ML-based approach to predict increased propensity for VOC crises in SCD. The promising results suggest the future possibility of embedding an improved version of this model in a low-cost wearable system to assist clinicians in managing long-term therapy for SCD patients.
Keywords
