Original title: Bridging the Gap in ECG-Based Emotion Recognition: A Unified Evaluation of Deep Learning Models
Timothy C Sweeney-Fanelli, Ajan Ahmed, Masudul Imtiaz · September 2026
Scientists want computers to recognize human emotions by reading heartbeats through electrocardiograms, or ECGs. However, comparing different artificial intelligence models is tricky because researchers test them with different methods and single, uniform datasets. To solve this problem, the authors built two free software tools to standardize testing and combine data. They used one tool to merge three public datasets into one large pool. This mix included different heart sensors, recording conditions, and diverse groups of people, forcing the models to face real variety. Next, the researchers used their benchmarking tool to test three popular deep learning models alongside two convolutional neural network baselines using a ten-fold testing process. They discovered important trade-offs between a model's complexity and its classification accuracy, creating a shared benchmark for future emotion recognition research.
Imagine baking the same cake recipe in three different kitchens. One oven runs hot, another has different pan sizes, and the third uses different flour, so a perfect result in one kitchen tells you nothing about the others. You cannot tell whether a bad cake means the recipe is wrong or the kitchen is different. Here, the recipe is the emotion-recognition program and the three kitchens are three public heart-signal datasets. The researchers build shared measuring tools and a fixed baking procedure so every recipe is judged under identical conditions.
Emotion recognition from ECGs could let devices sense stress or mood, but results across studies were impossible to compare. This paper standardizes how models are trained and tested, and merges three public datasets so a model must work across different sensors, conditions, and people. That makes published accuracy numbers meaningful and gives researchers a shared starting point.
The abstract reports a benchmark and trade-offs but does not state which model performed best, absolute accuracy numbers, or how well any model generalized. It also does not compare against all existing AER architectures — only three deep learning models and two CNN baselines. The merged dataset increases variability but still draws on only three public datasets, so results may not extend to other ECG collections or real-world settings.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Bridging the Gap in ECG-Based Emotion Recognition: A Unified Evaluation of Deep Learning Models — Timothy C Sweeney-Fanelli, Ajan Ahmed, Masudul Imtiaz
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.