Machine learning model separates depressed and non-depressed young patients with strong test performance
Depression in adolescents and young adults often presents without a dramatic clinical signal. Fatigue, sleep disruption, school decline, irritability, and somatic complaints can look nonspecific long before a clear psychiatric picture emerges. In that setting, the attraction of a data-driven diagnostic model is straightforward: if ordinary clinical review misses the early pattern, laboratory and demographic signals may help close the gap.
A retrospective cohort study involving 600 patients aged 10 to 26 years tested that idea directly. The sample included 350 non-depressed patients and 250 patients with depression, with data collected between 2019 and 2025. Investigators built a machine learning model for depression diagnosis and compared performance across approaches, then used SHAP plots to interpret the model and visualize feature contribution. The strongest performer was a gradient boosting machine, which reached a threshold of 0.591 with an accuracy of 0.878, specificity of 0.941, and precision of 0.913. Sensitivity reached 0.797 and the F1 score was 0.851.
Specificity at that level matters in a psychiatric diagnostic workflow. A model that flags too many healthy young people as depressed can add noise, anxiety, and unnecessary follow-up. Here, the high specificity suggests the GBM was relatively disciplined in excluding non-depressed patients, while the more moderate sensitivity shows it still missed some depressed cases. In clinical terms, the tool looks better suited as a support layer than as a stand-alone diagnostic gatekeeper.
S100β, platelet measures, and LDH ratios carried most of the signal
Feature selection was narrowed with Boruta, which left nine variables for model construction: S100β, platelet count, lactate dehydrogenase, white blood cell count, creatinine kinase-MB, gender, and three derived ratios — LDH/PLT, LDH/WBC, and LDH/HGB. The most significant features were S100β, platelet count, and LDH/PLT. Those variables point toward a broader biological footprint than mood symptoms alone.
S100β is often discussed in relation to nervous system injury and glial activity, so its prominence raises the possibility that neurobiological stress pathways may leave measurable traces in peripheral blood. Platelet measures and LDH-based ratios may reflect inflammatory or metabolic shifts that accompany depression, especially in younger patients whose presentations are not yet fully fixed into a chronic psychiatric pattern. None of these markers diagnose depression on their own, but their combined behavior inside a machine learning framework suggests that depression may be detectable as a multivariable physiologic signature rather than a single abnormal lab result.
The inclusion of gender also reflects how depression risk and expression vary across young populations. A model that ignores sex-related differences can lose calibration quickly, especially when the target population spans adolescence and early adulthood, a period marked by hormonal change, neurodevelopment, and shifting social stressors.
Calibration and decision-curve analysis supported clinical usability
Performance alone does not establish whether a prediction model will work at the bedside. Calibration and decision-curve analysis were used to assess whether the GBM’s probabilities tracked reality and whether the model provided net clinical benefit across decision thresholds. Both assessments favored the GBM, and calibration was reported as consistent with the model’s predictions.
This is the part that matters for real-world use. A diagnostic model can look impressive on paper and still fail in practice if its risk estimates are poorly aligned with actual disease frequency. The calibration signal here suggests the model’s predicted probabilities were not wildly detached from observed outcomes. Decision-curve analysis adds another layer by asking whether using the model would improve decisions compared with treating everyone or no one as potentially depressed.
In a young population, that matters because the cost of delay can be substantial. Depression during these years can interfere with education, relationships, physical development, and long-term mental health trajectory. A structured model built from routinely available variables could help direct attention to patients who warrant a more careful psychiatric evaluation, especially in settings where specialist access is limited.
Why this approach may fit young patients better than symptom screening alone
Young patients do not always describe depressive symptoms in a neat, textbook way. Somatic symptoms, irritability, fatigue, and academic problems may dominate the presentation, while verbalized sadness may be absent or minimized. A model that integrates laboratory data with basic demographic information may therefore catch patterns that a brief clinical encounter misses.
The biological logic is plausible even if the model remains preliminary. Depression is not only a disorder of mood; it is also associated with immune, metabolic, and neurobiological changes that can be reflected in peripheral measures. Ratios such as LDH/PLT or LDH/WBC may function as compressed summaries of systemic stress, while S100β may capture a more direct nervous system-related signal. Machine learning is particularly suited to this kind of signal mixing, because it can weigh small interactions across variables that would be easy to overlook in a conventional regression approach.
Still, the model’s performance should be interpreted with restraint. A retrospective design cannot establish causality, and a single-center data structure may limit portability to other hospitals, regions, and laboratory platforms. The study also does not show whether the model improves clinician decision-making, reduces missed diagnoses, or changes patient outcomes in practice. Those are the questions that matter before any diagnostic algorithm can move beyond internal development.
What the current evidence supports — and what it does not
This study provides preliminary evidence that machine learning can help identify depression in patients aged 10 to 26 years using a small set of clinical and laboratory variables. Among the tested approaches, the GBM delivered the best balance of discrimination, specificity, and calibration. The result is encouraging for diagnostic support in young populations, where depression is often subtle and easily confused with other complaints.
The work stops short of proving readiness for routine deployment. External validation, prospective testing, and comparison with standard clinical screening remain necessary before the model can be considered more than an investigational tool. The study also leaves open whether these biomarkers remain stable across different settings, ethnic groups, or stages of depressive illness. For now, the main contribution is a usable framework: depression in young people may leave a detectable pattern in blood-based and demographic data, and machine learning can organize that pattern into a model with meaningful diagnostic performance.