Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Prototype-Based Learning for Healthcare: A Demonstration of Interpretable AI

Ashish Rana, Ammar Shaker, Sascha Saralajew, Takashi Suzuki, Kosuke Yasuda, Shintaro Kato, Toshikazu Wada, Toshiyuki Fujikawa, Toru Kikutsuji

TL;DR

The paper addresses the need for interpretable and verifiable AI in personalized preventive healthcare. It introduces ProtoPal, a two-mode prototype-based framework that trains prototypes with Generalized Learning Vector Quantization and uses denoising autoencoders to simulate health trajectories, anchored by an 'ideal healthy self' representation. The online module provides a risk explainer and a greedy health planner to propose gradual lifestyle changes and visualize their impact via digital twins. Experiments on Kurashiki Central Hospital data show competitive AUC gains across 17 ICD-10 diseases compared with a Cox model, and the system yields interpretable patterns and actionable interventions suitable for clinical workflows.

Abstract

Despite recent advances in machine learning and explainable AI, a gap remains in personalized preventive healthcare: predictions, interventions, and recommendations should be both understandable and verifiable for all stakeholders in the healthcare sector. We present a demonstration of how prototype-based learning can address these needs. Our proposed framework, ProtoPal, features both front- and back-end modes; it achieves superior quantitative performance while also providing an intuitive presentation of interventions and their simulated outcomes.

Prototype-Based Learning for Healthcare: A Demonstration of Interpretable AI

TL;DR

The paper addresses the need for interpretable and verifiable AI in personalized preventive healthcare. It introduces ProtoPal, a two-mode prototype-based framework that trains prototypes with Generalized Learning Vector Quantization and uses denoising autoencoders to simulate health trajectories, anchored by an 'ideal healthy self' representation. The online module provides a risk explainer and a greedy health planner to propose gradual lifestyle changes and visualize their impact via digital twins. Experiments on Kurashiki Central Hospital data show competitive AUC gains across 17 ICD-10 diseases compared with a Cox model, and the system yields interpretable patterns and actionable interventions suitable for clinical workflows.

Abstract

Despite recent advances in machine learning and explainable AI, a gap remains in personalized preventive healthcare: predictions, interventions, and recommendations should be both understandable and verifiable for all stakeholders in the healthcare sector. We present a demonstration of how prototype-based learning can address these needs. Our proposed framework, ProtoPal, features both front- and back-end modes; it achieves superior quantitative performance while also providing an intuitive presentation of interventions and their simulated outcomes.
Paper Structure (6 sections, 2 equations, 4 figures, 1 table)

This paper contains 6 sections, 2 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Kurashiki Central Hospital
  3. Framework Structure
  4. Offline Mode: Interpretable Prototype-based Learning
  5. Online Mode: Interactive Personalized Healthcare Tool
  6. Conclusion and Outlook

Figures (4)

  • Figure 1: The architecture of ProtoPal. Left: The offline mode, which includes training GTLVQ and fitting denoising autoencoders. Right: The online mode, which first features a risk explainer that predicts an individual’s risk score. And second, a health planner that recommends a trajectory of lifestyle changes and illustrates how simulated biological vitals evolve through a series of intermediate healthier digital twins.
  • Figure 2: The continuous dot plots with line legends demonstrate that the learnt models capture disease biology (top row) and assists in identifying reasoning behind correct or incorrect predictions (bottom row).
  • Figure 3: The risk explainer output for a given highly diseased old individual, where disease risk dashboard summarizes all the different disease risks amongst various diseases. Additionally, detailed analysis of Type-2 Diabetes Mellitus (E11) is provided as well with the assistance from healthy and diseased digital twins. These digital twins are respectively created from the closest healthy ($P_H$) and closest diseased prototypes ($P_D$) after wholly adopting the respective lifestyle and correspondingly simulating all the biological attributes. Finally, the detailed lifestyle choices and simulated biology comparison bar plots contrast the individual's attributes against diseased and healthy digital twins.
  • Figure 4: The health planner for the same highly diseased old individual provides personalized iterative healthy lifestyle recommendations. This is done by using the closest healthy prototype to identify the best possible lifestyle recommendation for best disease risk reduction at each step. Also, first improvements in relatively easier to follow healthy eating habits are recommended. And reduction in harder or addictive activities like alcohol consumption is recommended in later stages for the elderly individual.