Psychological resilience, the capacity to maintain or recover healthy functioning in the face of stress and adversity, has long been a central concern of developmental psychology. Children with low resilience-related capacity are at elevated risk for emotional and behavioral difficulties as they grow, and researchers have spent decades trying to identify which factors distinguish children who bend from those who break. What has often been missing, however, is a way to translate that knowledge into practical, scalable screening tools that schools can actually use. The new study set out to fill that gap by asking a deceptively simple question: can machine learning models, trained on ordinary questionnaire data, reliably flag students with relatively low resilience before problems escalate?

The researchers grounded their work in the Promotive and Protective Factors and Processes framework, a widely used model that organizes the influences on resilience into internal characteristics of the child and external resources in the environment. Eleven predictors were measured, spanning both categories, using self-report questionnaires completed by the students themselves and by one of their parents. The measures drew on well-validated instruments, including the Strengths and Difficulties Questionnaire, which captures emotional symptoms, hyperactivity, conduct problems, peer problems, and prosocial behavior, as well as scales assessing temperament, perceived social support, and school climate. This longitudinal design, with data collected at two time points, allowed the team to test whether factors measured early could predict resilience-related capacity months later, a stronger test than a single snapshot.

On the modeling side, the team cast a wide net. Eight algorithms were trained and compared: logistic regression, decision tree, random forest, k-nearest neighbors, support vector machine, AdaBoost, LightGBM, and CatBoost. The latter two belong to the gradient boosting family, methods that build powerful ensembles of weak learners and have dominated machine learning competitions in recent years. Each model was evaluated using ten-fold cross-validation in the training set, a procedure in which the data are repeatedly split so that every case serves as a test for a model trained on the others, reducing the risk that good performance is merely an artifact of one lucky split.

The verdict was striking. Logistic regression, the simplest and most transparent of the eight, achieved the best performance for the study’s screening-oriented goal, posting the highest area under the receiver operating characteristic curve at 0.827, along with the best recall of 0.743 and the best F1 score of 0.621. When the final model was applied to an independent test set it had never seen, it held up well, achieving an AUC of 0.829, accuracy of 0.730, precision of 0.500, recall of 0.729, and an F1 score of 0.593. In practical terms, the model correctly identified roughly three out of four children with low resilience-related capacity, though half of its positive flags were false alarms, a trade-off that may be acceptable in a screening context where a flag leads to supportive follow-up rather than a diagnosis.

The result echoes a growing body of evidence in psychology and forensic science that complex, high-capacity models do not automatically outperform well-specified simpler ones, particularly on tabular data of modest size. Gradient boosting methods can capture intricate nonlinear interactions, but they also risk overfitting when the number of informative features is limited and the sample, while respectable at over a thousand participants, is not enormous. Logistic regression, by contrast, imposes a simple linear structure that appears to match the underlying signal in this dataset. For schools considering such tools, the simplicity bonus is doubled: the model that predicts best is also the one whose inner workings are easiest to explain to teachers, parents, and ethics boards.

That explainability was not left to chance. The researchers supplemented traditional feature importance rankings with SHAP, or SHapley Additive exPlanations, a technique borrowed from cooperative game theory that assigns each feature a contribution value for every individual prediction. Rather than treating the model as a black box, SHAP analysis reveals how much each factor pushed a given child’s predicted risk up or down, and in which direction. This dual approach, combining global feature importance with case-level explanations, represents the current best practice in what has become known as explainable artificial intelligence, and it is what allowed the team to move beyond raw prediction scores to substantive psychological insight.

What emerged from those analyses was a clear hierarchy of predictors. The five most important were hyperactivity, negative affectivity, peer support, emotional symptoms, and prosocial behavior. Notably, four of the five are internal characteristics of the child or closely tied to the child’s interpersonal world, rather than broad environmental factors such as family socioeconomic status or school climate. The authors interpret this pattern as suggesting that relatively low resilience-related capacity among Chinese elementary school students under ordinary developmental pressures is particularly associated with internal vulnerabilities and lower levels of individual and interpersonal protective resources. In other words, the children most likely to struggle are those carrying temperamental and emotional burdens, and lacking the social scaffolding of supportive peers and their own habits of helpful, cooperative behavior.

Each of these top predictors points toward a concrete avenue for intervention. Hyperactivity and emotional symptoms are already frequent targets of school-based mental health programs, and the findings reinforce the value of identifying children with elevated scores on these dimensions early. Negative affectivity, a temperamental tendency toward distress, is harder to change but can inform how adults interpret and respond to a child’s reactions. Peer support and prosocial behavior, meanwhile, are genuinely modifiable: schools can cultivate inclusive classroom climates, structured cooperative activities, and opportunities for children to practice helping others, which prior research links to stronger well-being and resilience. The reciprocity is intuitive, since children who help others tend to build the friendships that later buffer stress.

The study’s limitations are worth keeping in view. The sample comprised Chinese elementary school students, and cultural context shapes both how resilience is expressed and how questionnaire items are answered, so generalization to other populations requires further testing. The predictors relied on self- and parent-report questionnaires rather than clinical interviews, and the model’s precision of 0.500 means any real-world deployment would need to treat predictions as prompts for human judgment, not verdicts. The researchers also note that data are available from the first author on reasonable request, supporting scrutiny of the work. Still, the study was approved by the Institutional Review Board of the School of Psychology at Capital Normal University, informed consent was obtained from all participating parents and children, and the work was supported by the National Natural Science Foundation of China and a Beijing Educational Science Planning grant.

For a field increasingly enamored with ever-larger models, the message of this research is refreshingly grounded. Predicting which children are at risk of low resilience does not require exotic algorithms; it requires good measures, a longitudinal design, honest validation on held-out data, and explanation tools that make the model’s reasoning legible. The fact that hyperactivity, negative affectivity, peer support, emotional symptoms, and prosocial behavior rose to the top gives schools a short, actionable checklist, and the fact that a transparent logistic regression did the predicting means the checklist comes with an interpretable engine behind it. As machine learning continues to seep into education and mental health, this study offers a template for doing it responsibly: prioritize the screening goal, compare simple and complex models on equal footing, and demand explanations before trusting the predictions.

Subject of Research: Explainable machine learning prediction of psychological resilience in Chinese elementary school students

Article Title: Explainable prediction of psychological resilience among Chinese elementary school students using machine learning

Article References: Li, T., Wang, M., & Li, Y. (2026). Explainable prediction of psychological resilience among Chinese elementary school students using machine learning. Current Psychology, 45(20), Article 1597. https://doi.org/10.1007/s12144-026-10158-w

Image Credits: AI Generated

DOI: 10.1007/s12144-026-10158-w

Keywords: psychological resilience, machine learning, logistic regression, SHAP, elementary school students, screening, hyperactivity, peer support, emotional symptoms, prosocial behavior, longitudinal study, child mental health