Towards Reducing Diagnostic Errors with Interpretable Risk Prediction

TL;DR

This work proposes a Neural Additive Model to make predictions backed by evidence with individualized risk estimates at time-points where clinicians are still uncertain, aiming to specifically mitigate delays in diagnosis and errors stemming from an incomplete differential.

Abstract

Many diagnostic errors occur because clinicians cannot easily access relevant information in patient Electronic Health Records (EHRs). In this work we propose a method to use LLMs to identify pieces of evidence in patient EHR data that indicate increased or decreased risk of specific diagnoses; our ultimate aim is to increase access to evidence and reduce diagnostic errors. In particular, we propose a Neural Additive Model to make predictions backed by evidence with individualized risk estimates at time-points where clinicians are still uncertain, aiming to specifically mitigate delays in diagnosis and errors stemming from an incomplete differential. To train such a model, it is necessary to infer temporally fine-grained retrospective labels of eventual "true" diagnoses. We do so with LLMs, to ensure that the input text is from before a confident diagnosis can be made. We use an LLM to retrieve an initial pool of evidence, but then refine this set of evidence according to correlations learned by the model. We conduct an in-depth evaluation of the usefulness of our approach by simulating how it might be used by a clinician to decide between a pre-defined list of differential diagnoses.

Figures (13)

• Figure 1: Inherently "interpretable" approaches to prediction. Typically, 'interpretable' models trade off between the expressiveness of intermediate representations and the faithfulness of the resulting interpretability to the models' true mechanisms. Our approach (D) manages to use very expressive intermediate representations in the form of abstractive natural language evidence while still maintaining true transparency during aggregation of this evidence. See Table \ref{['tab:inherently-interpretable-approaches']} for more details.
• Figure 2: Explainable Risk Prediction and Training. An overview of our approach. Left: We retrieve evidence snippets from past notes with an LLM for predefined queries posed by a clinician. Then we use our risk prediction model to estimate risk of various diagnoses given each piece of evidence individually, and aggregate these scores. Right: We automatically extract diagnosis 'labels' from future reports with an LLM to use to train the risk predictor.
• Figure 3: Evidence Usefulness (the maximum score across conditions) for our approach and two ablations. "LLM Evidence+Confidence Sorting" uses model evidence, but sorts by (length-normalized) log probability instead of the log odds. "All EHR+Log Odds Sorting" does not use LLM evidence and instead takes the last 1000 sentences in the record as evidence.
• Figure 4: Seen vs. unseen evidence counts for all evidence that at least weakly correlates with a condition. Curiously, the LLM Evidence with Log Odds Sorting model has some hallucinated evidence that was seen by annotators. See section \ref{['section:results']} for a discussion.
• Figure 5: Synthetic label precision. For each confident diagnosis label extracted by the system, annotators check whether the diagnosis actually appears in the report (and is definitive), and subsequently if subjectively they believe that report is likely the first time the diagnosis was definitive based on the report language.