Assessment of PTSD in military personnel via machine learning based on physiological habituation in a virtual immersive environment

Figures

• 

  Fig. 1. Top-down view of the virtual immersive simulation. The participants were presented with this novel and real-world complex environment on a desktop screen. Their task was to search a scene for specific objects (e.g., all motorcycles or cars) while they were navigating the environment. The pacing of the search task was left to the participant, as was the navigation of the environment. Navigation cues (e.g., trail blazes) were provided to assist the participants and to provide a common perceptual target across all participants. Navigation was performed via standard game control keys (i.e., the W, A, S, and D keys), and the mouse was used to change the view or turn around. When the participant navigated near a navigation cue, a direction mark was displayed on the cue to indicate which specific direction the participant should be heading.

• 

  Fig. 2. Examples of the physiological signals used in this work. The HR was computed from an ECG signal captured at 2048 Hz. The phasic GSR was obtained from the raw GSR data collected at 2048 Hz via standard decomposition techniques (i.e., the convex optimization approach27). The pupil diameter was captured with the eye tracking sensor at 300 Hz. All three signals were downsampled at 32 Hz. The gray vertical lines correspond to the occurrence of an event during the simulation, and the red line corresponds to the event ‘GAMESTART’, i.e., at the beginning of the simulation.

• 

  Fig. 3. Schematic of how the habituation feature was computed from the physiological time series HR, phasic GSR, pupil diameter and event occurrence. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {Event_0}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {Event_{\textrm{last}}}$$\end{document} correspond to the first and last events of a given type, respectively, that occurred during the simulation.

• 

  Fig. 4. (a) Left: ROC curve of the prediction model. (b) Right: Confusion matrix (with normalized values in parentheses) of the model’s binary prediction for the test participants. With a binary accuracy of 80.95%, the model had a true negative rate (TNR) of 100% but a true positive rate (TPR) of 60%, meaning that 4 out of 10 participants with PTSD were not well detected. However, the ROC curve on the left indicates that this value could be increased to 80% at the cost of a slightly deteriorated false positive rate, meaning that more participants without PTSD would be predicted to have a PTSD. Depending on how important not missing any potential participants with PTSD is considered in the model, possibly optimal models could be fine-tuned with the desired TPR/TNR thresholds.

• 

  Fig. 5. SHAP explainability algorithm outputs for 3 different test participants. The example on top is for a participant who declared having a PTSD correctly predicted by the ML model. The middle example is for a participant without PTSD, which was also correctly predicted. Finally, the last example is for a participant who declared having a PTSD incorrectly detected as not having PTSD. The x-axis corresponds to the importance value determined via SHAP to the features on the y-axis. If the feature’s value is positive (in green), it supports the prediction of class 1 (PTSD), and if the feature’s value is negative (in red), it supports the prediction of class 0 (non-PTSD). The higher the absolute value of the feature is, the more important it is considered by the model to make its prediction. In these examples, the model’s prediction is always supported by several features, and their importance varies across the participants, which could be interpreted as an individual stress signature.

• 

  Fig. 6. (a) Top: Ranking of the cumulative absolute importance values of the features during the whole simulation for all participants. For each feature composed of an event/physiological marker pair, we summed the absolute importance value determined via SHAP for each test participant. The results provide an overview of the most important features for predicting PTSD or non-PTSD status across all participants and, therefore, which events are the most stressful, as well as which physiological features are the most related to these stressful stimuli. (b) Bottom: Ranking of the cumulative absolute importance values of the physiological features during the whole simulation for all participants. This result was obtained by grouping features of the same physiological category (e.g., max GSR, EBP, etc.). It provides an overview of which physiological features are the most predictive of PTSD or non-PTSD status across all participants.

• 

  Fig. 7. Individual habituation response as defined in Equation 2 versus the reported PCL-M score. A negative value on the x-axis means that the participant tends to have overall sensitization (i.e., negative habituation) to stress-inducing events during the scenario, whereas a positive value indicates a tendency to habituation to stress-inducing events. The orange line represents the linear regression of the scattered points in blue. These findings indicate that participants with low PCL-M scores tend to be more sensitive to stress stimulation. A participant was considered to have a PTSD when the PCL-M score was greater than 36.