The dataset for this work was sourced from two Institutional Review Board (IRB)-approved studies: (1) The Triheptanoin-clinical trial [27 ] (2) The Outcome measures and biomarkers development study [28 ]. The data were collected between January 2016 and December 2018—a three-year period. We used the body-worn patch BioStamp® (MC10 Inc., Cambridge, MA, USA) [29 ] to record ECG and three-axis acceleration from all the participants. While some ECG records were captured at a sampling rate of 125Hz, others were captured at a sampling rate of 250Hz. Concurrently, the three-axis acceleration records were captured at the sampling rates of 31.25Hz and 62.5Hz, respectively. These differences did not meaningfully influence the HRV and activity metrics we extracted [30 ]. We captured the ECG signal and the three-axis acceleration from the following four locations on the body: (1) Medial chest, (2) Left Hypochondrium, (3) Right Hypochondrium, and (4) Left Pectoralis. Per the protocols, all four locations were not used for all the participants, and only a subset of these locations was used for each participant. In conjunction to the signal data obtained from the biosensors, caretaker and physician surveys were conducted to obtain symptom severity for all 20 patients enrolled in the study. Specifically, the CGI-S scores were obtained through physician surveys to assign a binary label (low-severity vs. high-severity) for each patient-visit. A patient-visit was assigned to the low-severity category if the CGI-S ≤4 and was assigned to the high-severity category if the CGI-S >4. For each patient-visit we needed two consecutive days of signal data for the feature extraction. By applying this filter, we obtained a total of 32 patient-visits with two consecutive days of signal data and the associated CGI-S label. Among the 32 patient-visits, we had 18 high-severity visits corresponding to 10 unique patients and 14 low-severity visits corresponding to 11 unique patients. One patient had both low-severity and high-severity visits. We considered each patient-visit a data point and had a total of 32 data points with an associated binary label for model development and analysis. The racial, ethnic and age breakdown of the study subjects are illustrated in Fig 2A–2C respectively. The low and the high-severity groups had a similar split in patients with respect to the race, ethnicity and age. We had 8 patients each in low and high-severity groups whose race was “Caucasian” and 2 patients each in the two groups who’s race was “Asian”. One patient’s race was unknown. Further, we had 8 patients each in low and high-severity groups whose ethnicity was “Non-Hispanic” and 2 patients each in the two groups whose race was “Hispanic”. One patient’s race was unknown. Finally, the median age at enrollment in the low and high-severity groups were 8.5 and 8 years respectively.
Free full text: Click here
