Deep continual learning for medical call incidents text classification under the presence of dataset shifts

Pablo Ferri, Vincenzo Lomonaco, Lucia C. Passaro, Antonio Felix de Castro, Purificación Sánchez-Cuesta, Carlos Sáez, Juan M. García-Gómez

Abstract: Highlights•Dataset shifts arose within Valencian medical emergency data between 2009 and 2019.•The performance of deep models predicting event severity was affected by these shifts.•Incorporating continual learning strategies improved model performance over time.•Continual fine-tuning through time offered the best performance-efficiency trade-off.•Deep continual text models may add clinical value to the triage support process.