Assessing the Rigor of Machine Learning in Physiological Signal Processing Applications

In the dynamic field of biomedical engineering, the pervasive integration of machine learning into physiological signal processing serves various purposes, from diagnostics to Brain-Computer Interface (BCI) and Human-Machine Interface (HMI) using techniques such as Electroencephalography (EEG), Electromyography (EMG), Electrocardiography (ECG), and others. Nonetheless, the inherent scientific diversity within biomedical research often poses challenges, with practices sometimes misaligned with machine learning and standard statistical principles. This review analyzes 82 influential articles (2018-2023) from IEEE Xplore, aiming to identify weaknesses and assess overall rigor. It emphasizes the need for enhanced research quality and reproducibility. The key findings reveal that in over half of the articles, the ratio of female-to-male participants recruited for data collection is below 50% Additionally, nearly 30% of the studies involve fewer than 10 subjects in data collection, with only 7% providing justification for their sample size. Moreover, only about 34% of the articles provide access to their data, and a mere 26% report performance using a confusion matrix. These insights underscore critical areas for improvement, enhancing the robustness and transparency of applications in the physiological signal processing domain.