Comprehensive Summary
This article by Daida et al. discusses the future of artificial intelligence, specifically machine learning and deep learning, and its challenges of incorporating into clinical practice. This article in particular does not revolve around a study conducted by the author but rather summarizes the methods and implications of different models. Dr. Daida begins with an introduction of how AI is trained by using preexisting data sets and how supervised learning differs from unsupervised learning. He distinguishes between the commonly confused terms: AI, machine learning (ML), and deep learning (DL), where AI is the broad category consisting of the latter two. The topic then shifts from traditional learning methods to current areas of study and development. For example, high-frequency oscillations (HFOs) are known to correlate with seizure areas in brain tissue; when 100,000 time-frequency plots of oscillations were input to a DL model, it was able to detect and classify the HFOs with over 95% accuracy. Similarly, other models and techniques were discussed in their application to epilepsy research, including pinpointing the location of the seizure onset zone (SOZ) and minimizing “overfitting” data (an anomaly where AI performs well on training data but poorly on new cases).
Outcomes and Implications
Despite several successes, AI regulation plays a significant role in the delay of clinical implementation. Especially in a high-stakes field such as medicine, AI training models currently do not possess the universality to accurately detect EEG biomarkers such as HFOs in minority populations. The lack of standardized definitions and equitable representation of patient demographics without the guidance of practicing healthcare workers can skew machine learning algorithms into misinterpreting data. Instead, Dr. Daida proposes the idea that publicly accessible EEG databases such as OpenNeuro should be more widely used, given that the data stored in these platforms are consistently checked for accuracy. In the future, the hope is to be able to utilize DL models to identify areas of seizures without the need to capture one spontaneously, thus reducing the amount of time patients must remain hospitalized.