L-VISP: LSTM Visualization for Interpretable Symptom Prediction in Patient Cohorts

Area of Science:

• Oncology
• Computer Science
• Data Science

Background:

• Symptom modeling in head and neck cancer is complex due to heterogeneous patient data.
• Deep learning, particularly Long Short-Term Memory Networks (LSTMs), shows promise for patient risk prediction.
• A key challenge with LSTMs is their low interpretability, necessitating collaboration between data modelers and clinical experts.

Purpose of the Study:

• To introduce L-VISP, a novel human-machine solution for interpretable LSTM modeling in clinical research.
• To enhance the understanding and validation of deep learning models in oncology symptom analysis.
• To facilitate multidisciplinary collaboration between data scientists and clinicians.

Main Methods:

• Development of L-VISP, a visual analytics tool for LSTM modeling.
• Implementation of custom visual encodings to make multiple LSTM variants interpretable.
• Evaluation of L-VISP through collaboration with data modelers and a clinical oncologist.

Main Results:

• L-VISP provides interpretable insights into LSTM model operations and performance.
• The tool supports analysis from model understanding to clinical context interpretation.
• Multidisciplinary evaluation confirmed the utility of L-VISP in a clinical research setting.

Conclusions:

• L-VISP offers a viable solution for making deep learning models interpretable in clinical research.
• Visual analytics can bridge the gap between complex AI models and clinical validation.
• Human-machine collaboration is crucial for advancing AI applications in healthcare.