Comparison of machine learning and deep learning models in manual strength prediction using anthropometric variables

Objective. This study evaluated the predictive performance of machine learning and deep learning models in estimating manual strength in men and women using anthropometric variables. Methods. Anthropometric and strength data were collected from 382 participants from the economically active population of Campeche, Mexico. Predictive models implemented included linear regression, random forest, AdaBoost, extreme gradient boosting, TabNet, TabPFN and a custom convolutional neural network. Their performance was assessed using the mean absolute error, mean squared error and explained variance score. Additionally, SHAP (SHapley Additive exPlanations) analysis was conducted to interpret feature importance across models. Results. Deep learning models such as TabNet and TabPFN demonstrated superior prediction accuracy for torque strength, capturing complex non-linear interactions. Linear regression exhibited better generalization, particularly for grip strength prediction. SHAP analysis consistently identified palmar length and elbow-to-fingertip length as the most influential anthropometric predictors. Ensemble methods like random forest and AdaBoost performed well on training data but showed a tendency to overfit. Conclusions. Although advanced models enhanced performance in specific tasks, linear regression remained the most robust for generalization. Feature importance analysis confirmed the biomechanical relevance of the selected predictors. Future applications should balance model complexity with the need for interpretability, depending on ergonomic objectives.