Methods and Applications: Advancing Underlying Cause of Death Inference Through Wide and Deep Model

The surveillance of causes of death in populations is vital for public health decision-making (1). A crucial aspect of this surveillance is accurately completing death certificates. Challenges such as inadequate medical knowledge, unfamiliarity with cause-of-death inference rules, and limited practical experience can result in inconsistent quality of cause-of-death registration. Quality assessments have revealed that the accuracy of manually deduced underlying causes of death varies from 55% to 84% (2-3).

In recent years, there has been a focus on utilizing artificial intelligence (AI) techniques to automatically identify the primary causes of death. Despite deep neural network models achieving 97.8% accuracy in this task (4), their practical applications are restricted by performance limitations. This study employs the Wide and Deep framework to enhance the accuracy and stability of deep learning models, aiming to better predict the underlying causes of death and improve cause of death surveillance.

The data set utilized in our analysis comprises mortality records collected through nation-wide cause-of-death surveillance systems in Fujian Province, spanning from 2016 to 2022 (Supplementary Table S1). We employed the International Statistical Classification of Diseases and Related Health Problems, 10th Revision (ICD-10), to systematically categorize the causes of death into 25 broad classifications, while preserving the specificity of 4-digit codes. Any records with inconsistencies such as inaccuracies within the cause-of-death sequence, incorrect deduction of the primary cause of death, ICD-10 coding mistakes as identified by specialists at the Fujian Provincial CDC, or missing information regarding the deceased’s sex and age, were omitted from our study. Upon the elimination of redundant entries via cross-referencing identity numbers, names, and dates of demise, the final cohort entailed 403,547 individual death records. Identifying particulars were redacted, with the preservation of birth and death dates, sex, the complete ICD-10 codes for the causes of death, and the duration between disease onset and mortality. Analyses focused on the primary cause of death codes.

Age was divided into four groups: ≤14 years, 15–44 years, 45–64 years, and ≥65 years. Label encoding was applied to classify all ICD-10 codes, age groups, and genders into integer classes. The durations between illnesses were determined, with any missing durations recorded as null; the collected data was normalized using z-score normalization.

In light of considerable variation in the prevalence of different primary causes of death (Supplementary Table S2), we utilized a subsampling strategy for the over-represented categories C and I, whereby 40,000 cases were chosen at random. Conversely, categories with fewer instances underwent up sampling. We constructed a multi-directed graph using the cause-of-death sequences extracted from our data. This graph featured causes as nodes and incorporated both their sequential relationships and the temporal intervals between successive diseases as edge attributes. From each category, random samples of primary causes of death, along with other associated causes within their sequence, were selected. We used the graph to reconstruct cause-of-death sequences with lengths varying from one to four causes. To ensure uniformity across all categories classified by ICD-10 codes, we adjusted the sample size to 40,000 cases per category through oversampling, creating a research dataset comprising one million cases. We allocated 10% of the original dataset for each category to create a test set, ensuring that the number of samples from each category did not surpass 4,000 (Supplementary Table S2). The data remainder, after the test set extraction, served for model training and validation, applying a 5-fold cross-validation technique (Supplementary Figure S1). Notably, statistical analysis revealed significant disparities in both gender and age profiles between the test set and the training dataset (Supplementary Table S3).

The Wide and Deep model integrates a linear component, Wide, for memorization, and a neural network, Deep, for generalization. It incorporates a Convolutional Neural Networks (CNN) into the deep component to detect complex patterns in structured data while maintaining essential rule-based patterns. This model aims to accurately predict causes of death by balancing linear and non-linear interactions of features (Figure 1). It is well-suited for tasks with categorical and continuous data and has shown efficacy in recommendation systems (5).

Figure 1. 

Schematic diagram of the Wide and Deep model structure.

Abbreviation: ICD=International Classification of Diseases; CNN=Convolutional Neural Networks.

The research compared the performance of various models including XGBoost, CNN, Gated Recurrent Unit (GRU), Transformer, and GRU with an attention mechanism (Attention GRU) for the task of inferring the underlying cause of death. Python 3.11.5 and PyTorch 2.1.0 were used for analysis. The evaluation on the test set was based on metrics such as weighted accuracy, precision, recall, F1-Score, and area under the curve (AUC). The outcomes of the model tests were presented as means and standard deviations (x±s).