RetainVis: Visual Analytics with Interpretable and Interactive Recurrent Neural Networks on Electronic Medical Records

Abstract: We have recently seen many successful applications of recurrent neural networks (RNNs) on electronic medical records (EMRs), which contain histories of patients' diagnoses, medications, and other various events, in order to predict the current and future states of patients. Despite the strong performance of RNNs, it is often challenging for users to understand why the model makes a particular prediction. Such black-box nature of RNNs can impede its wide adoption in clinical practice. Furthermore, we have no established methods to interactively leverage users' domain expertise and prior knowledge as inputs for steering the model. Therefore, our design study aims to provide a visual analytics solution to increase interpretability and interactivity of RNNs via a joint effort of medical experts, artificial intelligence scientists, and visual analytics researchers. Following the iterative design process between the experts, we design, implement, and evaluate a visual analytics tool called RetainVis, which couples a newly improved, interpretable and interactive RNN-based model called RetainEX and visualizations for users' exploration of EMR data in the context of prediction tasks. Our study shows the effective use of RetainVis for gaining insights into how individual medical codes contribute to making risk predictions, using EMRs of patients with heart failure and cataract symptoms. Our study also demonstrates how we made substantial changes to the state-of-the-art RNN model called RETAIN in order to make use of temporal information and increase interactivity. This study will provide a useful guideline for researchers that aim to design an interpretable and interactive visual analytics tool for RNNs.

• The paper presents RetainEX, a novel bidirectional RNN that incorporates temporal attention and time intervals to enhance EMR predictions.
• It demonstrates improved predictive accuracy for conditions like heart failure, with AUC rising from 0.905 to 0.954.
• RetainVis offers an interactive platform enabling clinicians to perform what-if analyses for better model interpretability.

Analysis of RetainVis: Interpretability and Interactivity in RNN-Based EMR Analytics

The paper "RetainVis: Visual Analytics with Interpretable and Interactive Recurrent Neural Networks on Electronic Medical Records" presents a significant contribution to the domain of medical informatics, particularly in utilizing recurrent neural networks (RNNs) for predictive analytics on electronic medical records (EMRs). The authors have developed RetainEX, an interpretable RNN model, and RetainVis, a visual analytics tool that augments this model for enhanced user interaction and understanding.

Methodological Innovation

RetainEX introduces a novel approach to RNNs by integrating temporal attention mechanisms, enabling a fine-grained interpretability at both visit and code levels of patient data. This is achieved by employing bi-directional RNNs that process EMR data not only sequentially but in reverse order as well, allowing the model to capture temporal nuances between visits more effectively. The model also incorporates time intervals as additional input features, improving the granularity of predictions.

The RetainVis platform adeptly combines the power of RetainEX with visualization capabilities. This integration provides clinicians and data analysts with the ability to conduct what-if analyses, modifying input data such as medical codes and temporal intervals, and immediately observing the impact on prediction outcomes.

Empirical Evaluation

The study demonstrates substantial advancements in predictive accuracy and interpretability. RetainEX outperforms baseline models including GRU and RETAIN, as evidenced by strong improvements in AUC and Average Precision metrics across two significant health conditions: heart failure and cataract. For instance, the AUC for heart failure diagnosis risk prediction increased to 0.954 from 0.905 observed with the RETAIN model. This improvement highlights the efficacy of incorporating bidirectional processing and temporal information.

Visual Analytics and Interaction

RetainVis provides an intuitive user interface divided into coherent sections facilitating different analytical tasks. Users can examine patient cohorts, explore detailed records of individual patients, and interactively test hypotheses through a patient editor interface. These capabilities empower users to understand the underlying factors influencing prediction scores and thereby offer enhanced interpretive insights.

Practical and Theoretical Implications

Practically, the integration of RetainVis into clinical workflows can potentially refine the accuracy and reliability of predictive diagnostics. The approach offers a tool for clinicians to engage with predictive models actively, rather than passively accepting model outputs. Theoretically, this work sets a precedent for future research in the explainability of deep learning models applied in healthcare, suggesting a paradigm that balances performance with interpretability.

Future Directions

Further research could explore the scalability of RetainVis, assessing its application across broader datasets and varied medical conditions beyond heart failure and cataract. Additionally, validating the tool in real-world clinical settings can provide insights into its utility and refine its design based on user feedback. The potential to generalize the approach to other temporal and sequential data in different domains offers an expansive landscape for innovation.

In conclusion, RetainVis offers a pioneering step towards enhancing the interpretability and interactivity of predictive modeling in healthcare, aligning with the strategic objectives of explainable AI in sensitive fields such as medical diagnostics. This paper serves as an essential resource for researchers intent on evolving further applications of RNNs augmented with robust visualization frameworks.