An interpretable outcome prediction model based on electronic health records and hierarchical attention

Abstract

Outcome prediction aims to predict the future health condition of patients from Electronic Health Record (EHR) data. Because of the sequential characteristic of EHR data, recurrent neural network (RNN)‐based outcome prediction methods have achieved state‐of‐the‐art results. However, the major drawback of RNN‐based outcome prediction methods is lack of interpretability, which would lead to trust issues. Aiming at this problem, this paper proposes interpretable outcome prediction model with hierarchical attention (IoHAN), an interpretable outcome prediction model by leveraging attention mechanism. The main novelty of IoHAN is that it can pinpoint the fine‐grained influence on the final prediction result of each medical component by decomposing the attention weights hierarchically into hospital visits, medical variables, and interactions between medical variables. We evaluated IoHAN on MIMIC‐III, a large real‐world EHR data set. The experiment results demonstrate that IoHAN can achieve higher prediction accuracy than state‐of‐the‐art outcome prediction models. In addition, the hierarchical decomposed attention weights can interpret the prediction results in a more natural and understandable way.

Reliable prediction of childhood obesity using only routinely collected EHRs is possible

“…We investigated the risk factors identified by the model by analyzing which predictors most attribute to the model’s prediction. We used the attention scores [32, 33] obtained from the LSTM cells, as a way to determine which features are given more attention (importance) by the model to predict the output (more details are provided in Supplementary D.1).…”

A deep learning model for online doctor rating prediction

“…Self‐attention has been shown to perform better than sequence‐based models such as bidirectional gated recurrent unit (GRU) with attention (Bi‐GRU‐ATT) when dealing with a sparse dataset (Wang et al, 2021). Hierarchical attention, an improvement over the attention mechanism, has been applied in document classification (Yang et al, 2016), detecting financial fraud (Craja et al, 2020), health outcome prediction (Du et al, 2022), and sentiment analysis (Wang et al, 2021). Yang et al (2016) applied attention at two semantic hierarchies, namely, word and sentence level, but did not consider self‐attention and a mechanism to capture long‐term dependencies, unlike this work.…”

A hybrid adaptive approach for instance transfer learning with dynamic and imbalanced data

“…With the worldwide adoption of electronic health record (EHR) systems, machine learning has made great strides in the secondary use of EHR data toward more accurate clinical risk prediction 1–5 . These studies have a fundamental assumption that the data distributions of the training and test sets are the same, and thus prediction model development is typically a one‐time activity.…”

“…With the worldwide adoption of electronic health record (EHR) systems, machine learning has made great strides in the secondary use of EHR data toward more accurate clinical risk prediction. [1][2][3][4][5] These studies have a fundamental assumption that the data distributions of the training and test sets are the same, and thus prediction model development is typically a onetime activity. However, clinical practices such as patient care and hospital conditions can change over time, and disease prevalence and cause can also change over time 6 ; both cases would lead to changes in the data distributions, resulting in model performance drift.…”

“…Compared with full retraining, adapting the current model to a new environment requires a more complex learning paradigm, such as online learning 11 and lifelong learning, 12 both of which are prominent for dynamic environments or concept drift scenarios, while lifelong learning is superior to online learning in preserving old knowledge against being overwritten by new knowledge. 12 However, they are not suitable for EHR-based clinical prediction modeling because (1) they are designed to tackle streaming data (such as stock price, sensor, and other time-series data), that is, learn from a sequence of data instances one by one at each time and continually adapt the current model to the new environment, and (2) most EHRbased modeling studies are based on cross-sectional and longitudinal data of irregular time points.…”