A Machine Learning-Based Framework for Multi-Level Stress Detection and Personalised Intervention Using Psychometric Data

Abstract

Stress represents an important public health problem in the academic, corporate and digital worlds. Psychometric data have long been the primary means of measuring stress; they are quantitative measures of psychological attributes obtained using standardised, validated instruments. They are the Occupational Stress Index (OSI), the Perceived Stress Scale (PSS), the Depression Anxiety Stress Scales (DASS) and social media addiction scales (BSMAS, FoMO). These instruments operationalise subjective mental states as structured numerical feature sets amenable to computational analysis. But classical statistical techniques, such as ANOVA cannot be used to generate individual-level predictions or personalised recommendations from such data. This article is a conceptual and literature-based synthesis contribution: it proposes a structured machine learning (ML) pipeline for stress detection and classification across three levels low, moderate and high - designed around psychometric feature vectors derived from these validated instruments, but does not report new empirical model training or original experimental results. We discuss six of the commonly used ML algorithms in published literature and fit them into the proposed pipeline. The ML algorithms include Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), XGBoost, Artificial Neural Network (ANN) and K-Nearest Neighbours (KNN). We have included recommended pipeline stages for handling class imbalance through SMOTE and model interpretability through feature importance based on SHAP, based on earlier published results. The suggested framework, which is not a proven deployed system but rather a blueprint for future empirical validation, outlines a three-class stress prediction and a personalised intervention recommendation for decision support.