A study published in Discover Artificial Intelligence has tested whether machine learning can separate health care workers with high psychosocial resilience from those with low resilience, and the results point to a modest but usable signal. The research team, based at Universidad Tecnica Particular de Loja, drew on survey data collected across the United States during the COVID-19 pandemic and compared three different classification methods against one another.
The framing is a deliberate departure from much of the existing literature. Most research into the mental health of health care workers concentrates on negative outcomes such as burnout, emotional exhaustion and attrition. This study instead tried to predict a positive capacity: the ability to cope actively with stress and to recover from it. That shift in emphasis matters for how support programs might eventually be designed, because it asks what keeps people functioning rather than only what breaks them down.
Three algorithms, one clear winner
The team trained three classification models on the same set of survey responses: logistic regression, a random forest classifier and a support vector machine (SVM). Each model was asked to assign respondents to one of two groups, high resilience or low resilience, and the algorithms were then compared on how often they got that assignment right.
- Logistic regression: 75.6% accuracy, with an area under the ROC curve of 0.816
- Random forest: 72.6% accuracy
- Support vector machine: 70.8% accuracy
Logistic regression produced the strongest results. The area under the ROC curve is a measure of how well a model separates two classes across every possible decision threshold, and a value of 0.816 indicates reasonably good discrimination, though it falls well short of perfection. Notably, the simplest of the three methods outperformed the two more elaborate alternatives, a pattern that appears regularly in tabular survey data where relationships between variables are broadly linear and the sample is moderate in size.
Where the training data came from
Models need examples to learn from. The researchers used the How Right Now Mental Health & Coping dataset, collected by NORC at the University of Chicago between 2021 and 2022. It contains responses from 2,055 people in the United States, gathered during a stretch of the pandemic when hospitals and clinics were operating under sustained pressure, staff shortages and unusually high patient demand.
Building a resilience index
Before any model could be trained, the researchers needed something measurable to predict. They constructed a resilience index from four psychometric variables: resilience itself, the ability to bounce back, control and confidence. Those components were combined and then divided into two levels using the median as the cutoff point, producing a binary target of high versus low resilience for each respondent. Splitting at the median guarantees a balanced set of classes, which keeps accuracy figures comparable across models, though it also means the two groups are defined relative to each other rather than against an absolute clinical threshold.
What predicts low resilience
When the models were inspected to see which inputs carried the most weight, three factors dominated. Stress was the strongest contributor, with a weight of 0.182, followed by depression at 0.160 and anxiety at 0.145. Hopelessness and changes in sleep patterns came next, completing a cluster of variables that together signaled a greater likelihood of landing in the low-resilience group.
The ordering is intuitive, but it is still informative. It suggests that the accumulated burden of distress, rather than any single symptom in isolation, shapes whether a health care worker retains the psychological resources needed to keep coping. Screening approaches built on this kind of weighting could, in principle, flag the people whose combination of stress, low mood and anxiety places them at higher risk, before that risk translates into burnout or departure from the profession.
Coping strategies appear to stack up
Coping behaviors entered the model with lower individual weights, yet they were consistently associated with greater resilience. Seeking social support, engaging in hobbies, prayer and meditation all pointed in the same direction, suggesting that these habits accompany higher resilience even if they contribute less to the prediction than distress-related variables do.
More striking was how those behaviors combined. The researchers found that the effect appeared to be cumulative: respondents who reported using several coping strategies at once had, on average, higher resilience scores than those who relied on just one. That pattern hints that breadth may matter as much as intensity when it comes to protective habits. It is also the kind of finding that a single snapshot survey cannot fully explain, since a cross-sectional design cannot establish whether the strategies produced the higher scores or whether already-resilient people were simply more likely to adopt them.
Limits of the evidence
The authors are candid about the boundaries of their results. The data capture a single point in time rather than tracking individuals over months or years, so the study cannot show how resilience rises or falls, or which way the relationship with coping behaviors runs. All measurements rest on participants' own reports, which can be shaped by mood, memory and social expectations. And the sample is drawn entirely from the United States.
Those constraints limit how far the conclusions can be extended to other health care systems, including those in Latin America, where working conditions, staffing models, pay structures and cultural attitudes toward mental health may all differ. A model trained on one country's respondents is not automatically transferable to another, even when the underlying psychological concepts are similar.
The proposed next steps
The research team lays out a clear path forward. First, validate the model with health care workers in Latin America to test whether the same predictors hold in a different labor market and health system context. Second, enrich the input data with additional measures, such as sleep data or physiological indicators, which could complement self-reported responses and reduce the study's reliance on subjective assessment.
If that validation succeeds, the approach could support something more practical than a research finding: a way to identify which staff members might benefit most from support, and which combinations of coping strategies appear to protect them. The authors stop short of claiming the model is ready for deployment, and the accuracy figures justify that caution.
Why the result matters
Health care workforces have absorbed enormous pressure in recent years, and the decision to model resilience rather than only deficit is a meaningful one. An accuracy rate of 75.6% is not a diagnostic tool, and the researchers do not present it as one. But it does suggest that routinely collected survey responses contain a learnable pattern, one that a straightforward logistic regression can read at least as well as heavier machine learning alternatives. For health systems weighing how to direct scarce mental health resources, that is a signal worth validating further.
This article is based on reporting by Medical Xpress. Read the original article.
Originally published on medicalxpress.com
