A Data-Centric Approach To Generate Faithful and High Quality Patient Summaries with Large Language Models

Stefan Hegselmann, Zejiang Shen, Florian Gierse, Monica Agrawal, David Sontag, Xiaoyi Jiang
Proceedings of the fifth Conference on Health, Inference, and Learning, PMLR 248:339-379, 2024.

Abstract

Patients often face difficulties in understanding their hospitalizations, while healthcare workers have limited resources to provide explanations. In this work, we investigate the potential of large language models to generate patient summaries based on doctors’ notes and study the effect of training data on the faithfulness and quality of the generated summaries. To this end, we release (i) a rigorous labeling protocol for errors in medical texts and (ii) a publicly available dataset of annotated hallucinations in 100 doctor-written and 100 generated summaries. We show that fine-tuning on hallucination-free data effectively reduces hallucinations from 2.60 to 1.55 per summary for Llama 2, while preserving relevant information. We observe a similar effect on GPT-4 (0.70 to 0.40), when the few-shot examples are hallucination-free. We also conduct a qualitative evaluation using hallucination-free and improved training data. We find that common quantitative metrics do not correlate well with faithfulness and quality. Finally, we test GPT-4 for automatic hallucination detection, which clearly outperforms common baselines.

Cite this Paper

BibTeX
@InProceedings{pmlr-v248-hegselmann24a, title = {A Data-Centric Approach To Generate Faithful and High Quality Patient Summaries with Large Language Models}, author = {Hegselmann, Stefan and Shen, Zejiang and Gierse, Florian and Agrawal, Monica and Sontag, David and Jiang, Xiaoyi}, booktitle = {Proceedings of the fifth Conference on Health, Inference, and Learning}, pages = {339--379}, year = {2024}, editor = {Pollard, Tom and Choi, Edward and Singhal, Pankhuri and Hughes, Michael and Sizikova, Elena and Mortazavi, Bobak and Chen, Irene and Wang, Fei and Sarker, Tasmie and McDermott, Matthew and Ghassemi, Marzyeh}, volume = {248}, series = {Proceedings of Machine Learning Research}, month = {27--28 Jun}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v248/main/assets/hegselmann24a/hegselmann24a.pdf}, url = {https://proceedings.mlr.press/v248/hegselmann24a.html}, abstract = {Patients often face difficulties in understanding their hospitalizations, while healthcare workers have limited resources to provide explanations. In this work, we investigate the potential of large language models to generate patient summaries based on doctors’ notes and study the effect of training data on the faithfulness and quality of the generated summaries. To this end, we release (i) a rigorous labeling protocol for errors in medical texts and (ii) a publicly available dataset of annotated hallucinations in 100 doctor-written and 100 generated summaries. We show that fine-tuning on hallucination-free data effectively reduces hallucinations from 2.60 to 1.55 per summary for Llama 2, while preserving relevant information. We observe a similar effect on GPT-4 (0.70 to 0.40), when the few-shot examples are hallucination-free. We also conduct a qualitative evaluation using hallucination-free and improved training data. We find that common quantitative metrics do not correlate well with faithfulness and quality. Finally, we test GPT-4 for automatic hallucination detection, which clearly outperforms common baselines.} }
Endnote
%0 Conference Paper %T A Data-Centric Approach To Generate Faithful and High Quality Patient Summaries with Large Language Models %A Stefan Hegselmann %A Zejiang Shen %A Florian Gierse %A Monica Agrawal %A David Sontag %A Xiaoyi Jiang %B Proceedings of the fifth Conference on Health, Inference, and Learning %C Proceedings of Machine Learning Research %D 2024 %E Tom Pollard %E Edward Choi %E Pankhuri Singhal %E Michael Hughes %E Elena Sizikova %E Bobak Mortazavi %E Irene Chen %E Fei Wang %E Tasmie Sarker %E Matthew McDermott %E Marzyeh Ghassemi %F pmlr-v248-hegselmann24a %I PMLR %P 339--379 %U https://proceedings.mlr.press/v248/hegselmann24a.html %V 248 %X Patients often face difficulties in understanding their hospitalizations, while healthcare workers have limited resources to provide explanations. In this work, we investigate the potential of large language models to generate patient summaries based on doctors’ notes and study the effect of training data on the faithfulness and quality of the generated summaries. To this end, we release (i) a rigorous labeling protocol for errors in medical texts and (ii) a publicly available dataset of annotated hallucinations in 100 doctor-written and 100 generated summaries. We show that fine-tuning on hallucination-free data effectively reduces hallucinations from 2.60 to 1.55 per summary for Llama 2, while preserving relevant information. We observe a similar effect on GPT-4 (0.70 to 0.40), when the few-shot examples are hallucination-free. We also conduct a qualitative evaluation using hallucination-free and improved training data. We find that common quantitative metrics do not correlate well with faithfulness and quality. Finally, we test GPT-4 for automatic hallucination detection, which clearly outperforms common baselines.
APA
Hegselmann, S., Shen, Z., Gierse, F., Agrawal, M., Sontag, D. & Jiang, X.. (2024). A Data-Centric Approach To Generate Faithful and High Quality Patient Summaries with Large Language Models. Proceedings of the fifth Conference on Health, Inference, and Learning, in Proceedings of Machine Learning Research 248:339-379 Available from https://proceedings.mlr.press/v248/hegselmann24a.html.

Related Material