Explainable Stacked Ensemble Deep Learning (SEDL) Framework to Determine Cause of Death from Verbal Autopsies

Verbal autopsies (VA) are commonly used in Low- and Medium-Income Countries (LMIC) to determine cause of death (CoD) where death occurs outside clinical settings, with the most commonly used international gold standard being physician medical certification. Interviewers elicit information from relatives of the deceased, regarding circumstances and events that might have led to death. This information is stored in textual format as VA narratives. The narratives entail detailed information that can be used to determine CoD. However, this approach still remains a manual task that is costly, inconsistent, time-consuming and subjective (prone to errors), amongst many drawbacks. As such, this negatively affects the VA reporting process, despite it being vital for strengthening health priorities and informing civil registration systems. Therefore, this study seeks to close this gap by applying novel deep learning (DL) interpretable approaches for reviewing VA narratives and generate CoD prediction in a timely, easily interpretable, cost-effective and error-free way. We validate our DL models using optimisation and performance accuracy machine learning (ML) curves as a function of training samples. We report on validation with training set accuracy (LSTM = 76.11%, CNN = 76.35%, and SEDL = 82.1%), validation accuracy (LSTM = 67.05%, CNN = 66.16%, and SEDL = 82%) and test set accuracy (LSTM = 67%, CNN = 66.2%, and SEDL = 82%) for our models. Furthermore, we also present Local Interpretable Model-agnostic Explanations (LIME) for ease of interpretability of the results, thereby building trust in the use of machines in healthcare. We presented robust deep learning methods to determine CoD from VAs, with the stacked ensemble deep learning (SEDL) approaches performing optimally and better than Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN). Our empirical results suggest that ensemble DL methods may be integrated in the CoD process to help experts get to a diagnosis. Ultimately, this will reduce the turnaround time needed by physicians to go through the narratives in order to be able to give an appropriate diagnosis, cut costs and minimise errors. This study was limited by the number of samples needed for training our models and the high levels of lexical variability in the words used in our textual information.