Classical Models Matched Transformers in Daily Psychological Distress Prediction
A prospective cohort study from the University of Utah suggests that compact machine learning models may be just as useful as transformers for forecasting daily psychological distress from passive smartphone sensing and brief self-report. Investigators compared four classical algorithms with four transformer-based approaches in 169 community-dwelling adults followed for 7 days, using accelerometer, GPS, and power-state data paired with end-of-day Brief Symptom Inventory-18 assessments. The central question was not simply whether distress could be predicted, but whether more computationally demanding architectures offered any real advantage over smaller models that are easier to train, deploy, and scale.
The answer was largely no.
Across 1,062 person-days, 11.2% met the case definition for significant distress, defined as a sex-specific BSI-18 Global Severity Index T-score of at least 63. Held-out evaluation included 210 records with 25 cases, and participant-level bootstrap confidence intervals were used to temper overconfidence in the performance estimates. The best discrimination came from the support vector machine, with an area under the receiver operating characteristic curve of 0.84, essentially tied with the multimodal fusion transformer at 0.84. Extreme gradient boosting followed close behind at 0.82. Sensor-only transformers, by contrast, performed at chance, with AUROCs between 0.49 and 0.50.
Sensor Streams Alone Were Not Enough
The gap between multimodal and sensor-only transformer performance is clinically instructive. Passive smartphone signals are attractive because they are continuous and low burden, yet they can be noisy proxies for internal state. GPS, motion, and charging behavior may capture routine disruption, movement restriction, or altered daily structure, but they do not directly encode subjective symptom burden. When the model relied on sensors alone, it appears to have lost the contextual information needed to separate ordinary fluctuation from psychologically meaningful distress.
Baseline survey items, especially those related to functional impairment, carried the greatest feature importance. In practical terms, self-report remained central even in a study designed to test digital phenotyping. The result suggests that static psychosocial data may anchor the signal, while sensor data provide only a modest layer of temporal context around it.
At the default classification threshold of 0.50, elastic net logistic regression emerged as the best classifier, with sensitivity of 0.56 and positive predictive value of 0.64. When the threshold moved near the outcome prevalence, to 0.12, the support vector machine detected all cases and reached a negative predictive value of 1.00, but positive predictive values remained low across all models, ranging from 0.13 to 0.26. In a clinical triage setting, that tradeoff matters: a model that rarely misses a distressed day may still generate many false alarms, especially when the underlying event rate is low.
Threshold Choice Shapes Bedside Utility
Model ranking by AUROC alone obscures the decision problem clinicians actually face. A classifier that performs well in aggregate can still be inconvenient or inefficient if it floods a care team with low-probability alerts. Decision-curve analysis addressed that issue directly and favored most models over treat-all and treat-none strategies across threshold probabilities between 0.05 and 0.25. Those thresholds are not abstract statistical artifacts; they represent the range where a team might reasonably decide whether to trigger outreach, prompt a brief assessment, or escalate support.
This is where the study becomes more than a competition between algorithms. Passive sensing may work best as a triage layer rather than a standalone diagnostic tool. If a model can reliably rule out distress on low-risk days, it could reduce unnecessary check-ins and concentrate attention on the minority of days when support is more likely to be needed. The current data suggest the strongest use case may be a hybrid workflow in which digital signals supplement, rather than replace, routine psychosocial assessment.
The computational burden also favors restraint. Classical models performed as well as transformers while remaining far less resource-intensive. In settings with limited infrastructure, that difference influences whether a tool can be maintained, updated, and audited over time. Bigger architectures may look appealing on paper, but if the predictive gain is absent, the extra complexity becomes difficult to justify.
Why Functional Impairment May Outperform Raw Sensor Noise
Functional impairment likely captured more of the latent clinical state than the passive streams did on their own. Distress is not merely an alteration in movement or location; it often presents as diminished role function, disrupted routines, and the subjective burden of coping. A brief survey item can sometimes summarize that state more efficiently than a week of motion and location traces. The fact that baseline psychosocial measures dominated feature importance suggests that the most informative signal may still come from what patients say about their own functioning.
That finding does not diminish the value of smartphone sensing. Instead, it defines its role more narrowly. Digital phenotyping may be best suited to refining risk estimates within a broader psychosocial framework, especially when repeated self-report is feasible and acceptable. In populations where daily questionnaires are unrealistic, passive sensing may still carry value, but these data alone may not be rich enough to resolve distress with high confidence.
The study was prospective and participant-level held-out testing strengthens the internal validity of the estimates, yet the preprint remains limited by sample size, a 7-day observation window, and the fact that it has not undergone peer review. External validation will have to show whether these performance patterns hold in different populations, across longer monitoring periods, and in real-world clinical workflows where adherence, missingness, and intervention response all complicate prediction.
For now, the main clinical message is straightforward: when daily psychological distress is the target, smaller classical models can keep pace with transformers, and the decisive signal may still come from self-report rather than from sensors alone. The remaining question is not whether smartphone data are usable, but which combinations of patient-reported and passive features produce reliable, actionable triage without overwhelming care teams with false positives.