Authors :
Presenting Author: Michele Jackson, BA – Boston Children's Hospital
Jianbin Tang, PhD – IBM research – IBM Australia, Melbourne, VIC, Australia; Xiaofan Wang, MD PhD – Boston Children’s Hospital and Harvard Medical School, Boston, MA, USA; Rima El Atrache, MD – Boston Children’s Hospital and Harvard Medical School, Boston, MA, USA; Theodore Sheehan, BS – Boston Children’s Hospital and Harvard Medical School, Boston, MA, USA; Sarah Cantley, BS – Boston Children’s Hospital and Harvard Medical School, Boston, MA, USA; Shuang Yu, PhD – IBM Research – IBM Australia, Melbourne, VIC, Australia; Umar Asif, PhD – IBM Research – IBM Australia, Melbourne, VIC, Australia; Jeffrey Rogers, PhD – IBM Digital Health – Digital Health, IBM T.J. Watson Research Center, Yorktown Heights, NY, USA; Iven Mareels, PhD – IBM Australia, Melbourne, VIC, Australia; Stefan Harrer, PhD – IBM Australia, Melbourne, VIC, Australia and Digital Health Cooperative Research Centre, Sandringham, VIC, Australia; Tobias Loddenkemper, MD – Boston Children’s Hospital and Harvard Medical School, Boston, MA, USA
Rationale:
Large-scale and accurate data collection is vital for healthcare research, especially in deep learning-based studies. Here, we present a comprehensive data collection system to enable seizure monitoring with wearables and machine learning analyses.
Methods:
We designed a database (REDCap, Nashville, TN) with an automated data cleaning and annotation tool to collect clinical and wearable sensor (E3/E4: Empatica, Milan, Italy) and video-EEG data for patients who wore wearable sensors during long-term video-EEG monitoring on either wrists/ankles in the setting of a seizure detection and prediction study. Wearable sensors recorded autonomic measures and recordings were manually downloaded to a file storage server. Clinical and EEG data were extracted from electronic medical records. Board-certified epileptologists determined seizure onset, offset, and semiology from video-EEG per the ILAE 2017 classification.
Our database design (Figure 1) consisted of four instruments with repeating instruments for multiple enrolments: 1) Patient demographics: study code, first enrolment age and date, 2) Wearable recording session metadata: enrolment age and date, and exclusion data, 3) Wearable placements: device code, placement location, date/time, recording exclusion data, wearable and EEG recording synchronization timestamps and methods, 4) Seizure annotations: EEG start and end date/time, seizure onset, offset, and semiology.
Our data collection system (Figure 2) included three automated data cleaning and annotation tools[1] to ensure data consistency among four database instruments and wearable recording files. Tool 1 “Wearable Sensor Data Check”: a) checks if wearable sensor recordings exist on the server, b) checks if sensor recording timestamps overlap with placement records, c) checks if sensor recording time range overlaps with corresponding EEG time range, d) automatically fills synchronization variables in placement instrument, and e) calculates time drift between the wearable and EEG clock. Tool 2 “Seizure Annotation Check”: a) checks for seizure instrument completion and labels each seizure according to analysis criteria, b) checks if seizure time/date annotations overlap with EEG time range, and c) checks for overlapped seizures. Tool 3 “Labelled Data Generation”: a) detects and removes wearable sensor recording periods where wearables were not worn and b) generates annotated wearable data ready for machine learning analyses.
Results:
The tool displays indicative errors and warning messages to help clinicians locate data errors. It improved data collection efficiency and enabled us to clean and annotate instrument data for 1252 wearable recordings of 450 patients with 501 enrolments. For patients with seizures, tool implementation resulted in a cleaned data set of 166 patients with 900 seizures (13254 hours of sensor data) for 40 seizure types.
Conclusions: The database design and automated data cleaning tool enabled us to collect and analyze data efficiently and ensure high-quality data for machine learning analyses. The system can be used to further enrich clinical and wearable datasets and foster impactful research for seizure monitoring.
Funding:
IBM Research and the Epilepsy Research Fund