Decoding accelerometry for classification and prediction of critically ill patients with severe brain injury

Our goal is to explore quantitative motor features in critically ill patients with severe brain injury (SBI). We hypothesized that computational decoding of these features would yield information on underlying neurological states and outcomes. Using wearable microsensors placed on all extremities, we recorded a median 24.1 (IQR: 22.8–25.1) hours of high-frequency accelerometry data per patient from a prospective cohort (n = 69) admitted to the ICU with SBI. Models were trained using time-, frequency-, and wavelet-domain features and levels of responsiveness and outcome as labels. The two primary tasks were detection of levels of responsiveness, assessed by motor sub-score of the Glasgow Coma Scale (GCSm), and prediction of functional outcome at discharge, measured with the Glasgow Outcome Scale–Extended (GOSE). Detection models achieved significant (AUC: 0.70 [95% CI: 0.53–0.85]) and consistent (observation windows: 12 min–9 h) discrimination of SBI patients capable of purposeful movement (GCSm > 4). Prediction models accurately discriminated patients of upper moderate disability or better (GOSE > 5) with 2–6 h of observation (AUC: 0.82 [95% CI: 0.75–0.90]). Results suggest that time series analysis of motor activity yields clinically relevant insights on underlying functional states and short-term outcomes in patients with SBI.


Fig. S1
CONSORT-style flow diagram for patient enrollment and follow-up 2

Fig. S2
Count histograms of accelerometry recording information 3

Fig. S3
Trajectories of motor component scores of the Glasgow Coma Scale (GCSm) of each study participant during ICU stay 4-7

Table S1
Count distributions of GCSm scores per observation window 8 Table S2 Discrimination of threshold-level GCSm detection models per observation window 9

Table S3
Count distributions of GOSE scores at hospital discharge per observation window 10

Table S4
Discrimination of threshold-level GOSE at hospital discharge prediction models per observation window 11

Table S5
Count distributions of GOSE scores at 12 months post discharge per observation window 13

Fig. S5
Discrimination performance of functional outcome at 12 months post discharge prediction models on validation sets 14

Table S6
Discrimination of threshold-level GOSE at 12 months post discharge prediction models per observation window 15

Fig. S6
Probability calibration of optimally discriminating motor function detection and functional outcome prediction models on validation sets 16-17

Table S7
Probability calibration metrics of optimally discriminating models 18

Fig. S7
Mean motion feature trajectories in the six hours preceding GCSm evaluation, stratified by GCSm scores and bilateral sensor placement Percentages of missing, static, and dynamic accelerometry data by time of day of recording and sensor placement

Recruitment criteria at time of study enrollment:
◆ Admitted to Neurosciences Critical Care Unit (NCCU) ◆ Impaired consciousness as a result of acute injury to or illness in the brain ◆ At least 18 years old ◆ Presence of both arms and both legs and no injuries or lesions that may impair placement of accelerometers on either arm nor on either leg ◆ Not expected to die or have withdrawal of life-sustaining therapies, per attending physician, within 24 hours 1) Threshold-level detection of concurrent GCSm 2) Threshold-level prediction of GOSE at hospital discharge  S3. Trajectories of motor component scores of the Glasgow Coma Scale (GCSm) of each study participant during ICU stay. Shaded areas represent time ranges during which we recorded accelerometry from the corresponding patient and points mark the exact times of a GCSm evaluation.  Values in the five rightmost columns represent mean validation set area under the receiver operating characteristic curve (AUC) values with associated 95% confidence intervals in parentheses. Confidence intervals were derived using bias-corrected bootstrapping (1,000 resamples) and represent the variation across repeated cross-validation folds (5 repeats of 5 folds) and nine missing value imputations. Missing values designate insufficient diversity in endpoint labels to evaluate models of that observation window and threshold combination.   ( Values in the five rightmost columns represent mean validation set area under the receiver operating characteristic curve (AUC) values with associated 95% confidence intervals in parentheses. Confidence intervals were derived using bias-corrected bootstrapping (1,000 resamples) and represent the variation across repeated cross-validation folds (5 repeats of 5 folds) and nine missing value imputations. Missing values designate insufficient diversity in endpoint labels to evaluate models of that observation window and threshold combination. Precision recall curve of optimally discriminating model configuration of GOSE > 5 prediction at hospital discharge (Fig. 3a). Shaded areas represent 95% confidence intervals derived using bias-corrected bootstrapping (1,000 resamples) to represent the variation across repeated cross-validation folds (5 repeats of 5 folds) and nine missing value imputations. The values in the box represent the observation window of the optimally discriminating model as well as the mean average precision (with 95% confidence interval in parentheses). The horizontal dashed line represents the line of no detection power, equivalent to the proportion of the positive class (average precision = 0.02). b Density histograms of predicted probabilities for positive cases (upward) and negative cases (downward) of GOSE > 5 prediction at hospital discharge. n represents the number of unique observations pertaining to each case and the range of predicted probabilities is fixed on a narrow, near-zero range to demonstrate the low predicted probabilities of the model.    Values in the seven rightmost columns represent mean validation set area under the receiver operating characteristic curve (AUC) values with associated 95% confidence intervals in parentheses. Confidence intervals were derived using bias-corrected bootstrapping (1,000 resamples) and represent the variation across repeated cross-validation folds (5 repeats of 5 folds) and nine missing value imputations. Missing values designate insufficient diversity in endpoint labels to evaluate models of that observation window and threshold combination.   pertaining to the observation windows with the highest achieved AUC (a) per each detection threshold of the motor component score of the Glasgow Coma Scale (GCSm) as shown in Fig. 2a and (b) per each tested prediction threshold of the Glasgow Outcome Scale -Extended (GOSE) as shown in Fig. 3a.

Supplementary
Shaded areas represent 95% confidence intervals derived using bias-corrected bootstrapping (1,000 resamples) to represent the variation across repeated cross-validation folds (5 repeats of 5 folds) and nine missing value imputations. The distribution of predicted probabilities is shown at the bottom of the graphs, stratified by threshold-level endpoints. The values in each box represent the observation window achieving the highest AUC as well as the corresponding mean maximal absolute difference between observed and predicted probabilities of the endpoint (Emax) with 95% confidence interval in parentheses. The diagonal dashed line represents the line of perfect calibration.  Probability calibration metrics [mean (95% confidence interval)] corresponding to models trained on observation windows that maximize the area under the receiver operating characteristic curve (AUC) for each threshold (Fig. 2a, 3a, and Supplementary Fig. S4 online). Confidence intervals were derived using bias-corrected bootstrapping (1,000 resamples) and represent the variation across repeated crossvalidation folds (5 repeats of 5 folds) and nine missing value imputations. Acronyms: motor component score of the Glasgow Coma Scale (GCSm), Glasgow Outcome Scale -Extended (GOSE), and Integrated Calibration Index (ICI). *Count distribution of negative vs. positive cases with the proportion of the most represented case, equivalent to the no information rate, in parentheses. were removed prior to calculation of the mean values. Shaded areas represent the 95% confidence interval bootstrapped from 1,000 resamples to represent the variation across unique GCSm observations. The solid dark red line on the rightmost edge of each graph represents the time of GCSm evaluation. Feature type acronyms are decoded in Table 3.

Supplementary Figure S7
Supplementary Figure S8   Fig. S8. Correlation matrices of extracted motion features across different sensor placements. Each matrix represents a unique feature type and values in each cell of the matrices represent the mean Spearman's rank correlation coefficient (r) between two sensor placements across the study population (n = 69) with the associated 95% confidence interval (bootstrapped with 10,000 resamples) in parentheses. Sensor placement acronyms correspond to the right and left elbows (RE and LE), the right and left wrists (RW and LW), and the right and left ankles (RA and LE). Feature type acronyms are decoded in Table 3. Outliers, defined as values extending beyond two times the interquartile range above the third quartile, were removed from the plot. Means of numerical distributions per GCSm score were each compared against the compiled distribution mean of all GCSm scores using the Wilcoxon signed-rank test. Statistically significant differences are marked with asterisks (*p ≤ 0.05, **p ≤ 0.01, ***p ≤ 0.001, ****p ≤ 0.0001). Feature type acronyms are decoded in Table 3.

Wavelet-domain features
**** **** **** **** **** *** Supplementary Figure S10  Thus, the light grey shaded area represents the percentage of total missing data, the light cyan shaded area represents the percentage of total static activity, and the light green shaded area (barely visible) represents the percentage of total dynamic activity.  Recording duration is specified in hours:minutes:seconds. The total percentage of missing data, across all patients and all sensors, is 6.76% (excluding bed sensor: 5.91%).