Linguistically Communicating Uncertainty in Patient-Facing Risk Prediction Models

TL;DR

This work tackles the challenge of communicating uncertainty in patient-facing AI-based risk tools in healthcare, arguing that trustworthy risk communication requires aligning explanations with patients' mental models. It synthesizes five facets of uncertainty—Performance Metrics, Confidence, Precision vs. Confidence, Model Reasoning, and Unknown Knowns—and demonstrates the need for patient-centered interfaces that combine local explanations, global context, and dialogue-based interactions, illustrated with CHD and IVF examples. The authors propose measurement and visualization approaches (e.g., $p$-values, reliability diagrams) and advocate qualitative studies to evaluate how explanations affect patients' mental models and trust. The study outlines a concrete IVF-focused research plan (OPIS-based) to test and refine uncertainty communication strategies, with the goal of improving understandability and expectation management in patient-facing AI tools.

Abstract

This paper addresses the unique challenges associated with uncertainty quantification in AI models when applied to patient-facing contexts within healthcare. Unlike traditional eXplainable Artificial Intelligence (XAI) methods tailored for model developers or domain experts, additional considerations of communicating in natural language, its presentation and evaluating understandability are necessary. We identify the challenges in communication model performance, confidence, reasoning and unknown knowns using natural language in the context of risk prediction. We propose a design aimed at addressing these challenges, focusing on the specific application of in-vitro fertilisation outcome prediction.

Figures (5)

• Figure 1: A patients perspective of risk communication and model explanation for a CHD prediction model.
• Figure 2: A CART model for predicting the risk of CHD. (Cholesterol HDL ratio - Ratio of total cholesterol to HDL cholesterol). The number of data points corresponding to each node is denoted as samples. Confidence of leaf node prediction based on data distribution at the node is computed using the chi-square test, and the corresponding p-value is displayed.
• Figure 3: The decision path followed along a given DT for a particular patient input. The model predicts low risk following the decision path highlighted in Blue.
• Figure 4: The cumulative probability over 6 IVF cycle displayed as graph in the OPIS tool. Corresponding patient input - Age = 34; Years of infertility = 0; Number of eggs collected in first IVF cycle = 1; Type of embryo transfer = Stage 2 embryos transferred on day 2 or 3 ; Previous pregnancy = No; Tubal infertility = No; First cycle type = IVF; Embryos frozen in first cycle = Yes.
• Figure 5: A verbal mapping of confidence based on precision and Gini index for CHD risk prediction. Here the range of precision and confidence scores are limited based on the model from Figure \ref{['fig:DT']}.