Random Forests Outperform Deep Learning in Predicting Exam Scores from Vital Signs
Simple models beat complex neural networks on heart rate and skin temp data.
Researchers Lala Yamazaki and Ramchandra Rimal used machine learning to predict exam outcomes from physiological signals including electrodermal activity, heart rate, and skin temperature. They tested logistic regression, random forest, SVM, LSTM, GRU, and transformers. While deep learning models generally captured complex relationships, simpler models like random forests sometimes achieved superior performance while offering computational efficiency and interpretability. Transformers showed versatility comparable to LSTM and GRU. The study underscores the importance of balancing precision, efficiency, and interpretability in model selection.
- Random forest models achieved up to 89% F1-score, outperforming LSTMs and GRUs on physiological exam data.
- Simple models (logistic regression, random forest) offered full interpretability, allowing researchers to identify key stress signals.
- Transformers demonstrated strong performance comparable to recurrent networks, showing their versatility on numerical tabular data.
Why It Matters
Enables low-cost, interpretable AI systems to monitor student stress and improve academic outcomes using wearable sensors.