This Patient Looks Like That Patient: Prototypical Networks for Interpretable Diagnosis Prediction from Clinical Text

Abstract

The use of deep neural models for diagnosis prediction from clinical text has shown promising results. However, in clinical practice such models must not only be accurate, but provide doctors with interpretable and helpful results. We introduce ProtoPatient, a novel method based on prototypical networks and label-wise attention with both of these abilities. ProtoPatient makes predictions based on parts of the text that are similar to prototypical patients-providing justifications that doctors understand. We evaluate the model on two publicly available clinical datasets and show that it outperforms existing baselines. Quantitative and qualitative evaluations with medical doctors further demonstrate that the model provides valuable explanations for clinical decision support.

Proto-lm: A Prototypical Network-Based Framework for Built-in Interpretability in Large Language Models

“…Their focus is on learning one representative prototype per class while their performance is dependent on the size of the support set in few-shot learning scenarios. Due to these limitations, there have been relatively few works that utilize prototypical networks to provide interpretability for LLM's in NLP (Garcia-Olano et al, 2022;Das et al, 2022;Van Aken et al, 2022). and Hase et al (2019) use prototypical parts networks with multiple learned prototypes per class but only apply their methods to image classification tasks.…”

“…Our work is most closely related to Das et al (2022) and Van Aken et al (2022) in terms of the approach taken. However, our work differs in that the architecture in (Das et al, 2022) only utilizes a single negative prototype for binary classification, while proto-lm enables multi-class classification by using multiple prototypes for each class.…”

“…Additionally, we have extended the work of (Das et al, 2022) by implementing token-level attention to identify not only influential samples but also influential sections of text within each sample. Moreover, different from the single-prototype-as-asummary approach in Van Aken et al (2022), by learning multiple prototypes per class, proto-lm creates a prototypical space ( §4.1), where, unlike the embedding space of LLM's, each dimension is meaningful, specifically the distance to a learned prototype, and can be used to explain a decision. Our work is also similar to Friedrich et al (2021) in terms of our loss function design.…”

“…In order to help human users understand prototypes in natural language, we identify the closest training data sample for each prototype and project the prototype onto that sample. The projection of prototypes onto the nearest sample is a well-studied and established technique (Das et al, 2022;Van Aken et al, 2022;Hase et al, 2019). We compare the quality of our projections against the projections obtained via the training procedure of Proto-tex (Das et al, 2022) and the loss function used in Protoypical Network by measuring the average normalized distance between each prototype and their projected sample's embedding.…”