Abstract
Language features may reflect underlying cognitive and emotional processes following a traumatic event that portend clinical outcomes. The authors sought to determine whether language features from usual smartphone use were markers associated with concurrent posttraumatic symptoms and worsening or improving posttraumatic symptoms over time following a traumatic exposure. This investigation was a secondary analysis of the Advancing Understanding of RecOvery afteR traumA study, a longitudinal study of traumatic outcomes among survivors recruited from 33 emergency departments across the United States. Adverse posttraumatic sequelae were assessed over the six months following the initial traumatic exposure. Language features were extracted from usual smartphone use in a specialized app. Bivariate linear mixed models were used to identify and validate language features that are markers associated with posttraumatic symptoms. Participants were 1744 trauma survivors, with a mean age of 39 [SD = 13] years old, and 56% were female. Fourteen language features were associated with severity level of posttraumatic symptoms at specific timepoints (cross-sectional markers) and five features were associated with change in severity level of posttraumatic symptoms (longitudinal markers). References to the body and health or illness were predictive of worsening pain, somatic, and thinking/concentration/fatigue symptom severity over time. An increase in references to others was associated with improvement in somatic symptom severity over time and increases in expressions of causation or cognitive processes were associated with improvement in pain symptom severity over time. Language features derived from usual smartphone use can convey important information about health, functioning, and recovery following a traumatic event. Clinicians might utilize such information to determine who may experience a high symptom burden or risk of worsening posttraumatic symptoms.
Introduction
Following a traumatic event—such as a motor vehicle collision or physical assault—survivors often develop a range of Adverse Post-Traumatic Neuropsychiatric Sequelae (APNS). These sequelae encompass not only PTSD but also chronic pain, depression, sleep disruption, and cognitive fatigue. A significant challenge in post-traumatic care is the clinical "blind spot" that occurs once a patient leaves the Emergency Department (ED). Traditional follow-up relies on periodic self-reporting or re-traumatizing narratives. This study investigates whether passive, everyday language typed into a smartphone can serve as "digital biomarkers" to monitor recovery trajectories without requiring patients to explicitly recount their trauma.
Methods
This research utilized data from the AURORA study, a longitudinal investigation of 1,744 trauma survivors recruited from 33 EDs across the U.S. Participants installed the Mindstrong Discovery app on their Android devices, which passively collected unordered word lists and frequencies typed via the keyboard. This method preserved privacy by excluding context and word order.
Linguistic features were categorized using the Language Inquiry and Word Count (LIWC) tool. The researchers employed bivariate linear mixed models to identify cross-sectional markers (associated with current symptom severity) and longitudinal markers (associated with changes in severity over six months). Ground truth was established through regular "flash surveys" assessing ten APNS symptom domains.
Results
The study identified 14 cross-sectional and 5 longitudinal language markers. While median symptom scores generally trended downward over the 200-day observation period, individual recovery paths varied significantly.
A primary finding was the predictive utility of specific word categories. Increased references to the "body," "health," or "illness" were strong predictors of worsening pain and somatic symptoms, suggesting a narrowing of focus onto physical suffering. Conversely, increases in words related to "causation" (e.g., "because") and "cognitive processes" (e.g., "think") were associated with symptom improvement, indicating that intellectual processing of the trauma may facilitate recovery.
The predictive models demonstrated high accuracy for identifying patients at risk of worsening symptoms, with Positive Predictive Values (PPV) ranging from 0.76 to 0.82.
Discussion and Limitations
The results suggest that digital phenotyping via language use can disambiguate symptom domains; for example, "sadness" words were specific to anxiety, while "anger" words were more indicative of depression. However, several limitations remain. The "iOS Gap"—the exclusion of Apple users due to keyboard privacy restrictions—limits generalizability. Additionally, the act of taking daily symptom surveys may have influenced participants' subsequent language choices. Future work should integrate these linguistic markers with other digital phenotypes, such as sleep patterns and keystroke dynamics, to enhance clinical utility.
Conclusion
Language features derived from routine smartphone use provide valuable insights into a survivor's health and recovery trajectory following trauma. These digital biomarkers offer a non-invasive means for clinicians to identify individuals at high risk of worsening symptoms, enabling more timely and targeted interventions.
Summary updated