Table 1.
Paroxysmal Sympathetic Hyperactivity-Assessment Measure (PSH-AM).
Fig 1.
High resolution clinical feature scale (hrCFS).
Method illustration for calculation of hourly high-resolution clinical feature scale (hrCFS) based on 0.5 Hz vital sign trend data.
Fig 2.
Expert system-based episode detection method.
Flow diagram illustrating the derivation process for the expert system (ES) method of detecting paroxysmal sympathetic hyperactivity (PSH) episodes based on an expert clinician’s annotations of continuous vital sign (VS) data from a patient with PSH. Annotation is performed by the expert clinician and rule extraction by the data analyst.
Fig 3.
Machine learning model-based episode detection method.
Flowchart illustrating the derivation process for the support vector machine (SVM) model-based method of detecting paroxysmal sympathetic hyperactivity (PSH) episodes using an expert clinician’s annotations of continuous vital sign (VS) data from a patient with PSH. Annotation is performed by the expert clinician and rule extraction by the data analyst.
Table 2.
Included patient characteristics.
Table 3.
Summary of detected PSH episodes for the first 14 days.
Fig 4.
Face validity of automatic PSH quantification methods to distinguish between clinically defined PSH cases and controls.
The high-resolution Clinical Feature Scale (hrCFS, 4A) and aggregate burden score trends from the expert system (ES) method (4B) and the machine learning (ML) model (4C) during the first 14 days of hospitalization are displayed, stratified by PSH clinical diagnosis. Score trends are computed from a series of 12-hour non-overlapping windows over the course of 14 days. Mean scores for PSH cases and controls are displayed as solid red and blue lines, respectively, and shaded areas represent 95% confidence intervals.
Table 4.
Episode detection method performance.
Fig 5.
Illustrations of episode detection method performance.
Examples of detected episodes (highlighted in yellow) and clinician’s annotations (highlighted in gray) overlayed onto heart rate, in beats per minute (bpm) tracings. 5A illustrates good performance with only small timing differences between annotated and detected episodes, and the majority of episodes are overlapping. 5B provides an example of a low-precision period, defined as detected episodes that were not annotated by the clinician (false positives). 5C illustrates a low-recall scenario, where clinician-annotated episodes were not detected by our methods (false negatives).
Fig 6.
Framework for Human-Machine collaborative annotation and modeling.
The above figure illustrates our envisioned human-in-the-loop method for iterative improvement of our PSH episode detection tools.