Abstract
Analyzing patient data under current mechanical ventilation (MV) management processes is essential to understand MV consequences over time and to hypothesize improvements to care. However, progress is complicated by the complexity of lung-ventilator system (LVS) interactions, patient-care and patient-ventilator heterogeneity, and a lack of classification schemes for observable behavior. Ventilator waveform data originate from patient-ventilator interactions within the LVS while care processes manage both patients and ventilator settings. This study develops a computational pipeline to segment joint waveform and care settings timeseries data into phenotypes of the data generating process. The modular framework supports many methodological choices for representing waveform data and unsupervised clustering. The pipeline is generalizable although empirical output is data- and algorithm-dependent. Applied individually to 35 ARDS patients including 8 with COVID-19, a median of 8 phenotypes capture 97% of data using naive similarity assumptions on waveform and MV settings data. Individual's phenotypes organize around ventilator mode, PEEP, and tidal volume with additional delineation of waveform behaviors. However, dynamics are not solely driven by setting changes. Fewer than 10% of phenotype changes link to ventilator settings directly. Evaluation of phenotype heterogeneity reveals LVS dynamics that cannot be discretized into sub-phenotypes without additional data or alternate assumptions. Individual phenotypes may also be aggregated for use in scalable analysis, as behaviors in the 35 patient cohort comprise 16 cohort-scale LVS types. Further, output phenotypes compactly discretize the data for longitudinal analysis and may be optimized to resolve features of interest for specific applications.