Improving Large Language Models’ Summarization Accuracy by Adding Highlights to Discharge Notes: Comparative Evaluation
Notice bibliographique
Résumé
BACKGROUND: The American Medical Association recommends that electronic health record (EHR) notes, often dense and written in nuanced language, be made readable for patients and laypeople, a practice we refer to as the simplification of discharge notes. Our approach to achieving the simplification of discharge notes involves a process of incremental simplification steps to achieve the ideal note. In this paper, we present the first step of this process. Large language models (LLMs) have demonstrated considerable success in text summarization. Such LLM summaries represent the content of EHR notes in an easier-to-read language. However, LLM summaries can also introduce inaccuracies. OBJECTIVE: This study aims to test the hypothesis that summaries generated by LLMs from highlighted discharge notes will achieve increased accuracy compared to those generated from the original notes. For this purpose, we aim to prove a hypothesis that summaries generated by LLMs of discharge notes in which detailed information is highlighted are likely to be more accurate than summaries of the original notes. METHODS: To test our hypothesis, we randomly sampled 15 discharge notes from the MIMIC III database and highlighted their detailed information using an interface terminology we previously developed with machine learning. This interface terminology was curated to encompass detailed information from the discharge notes. The highlighted discharge notes distinguished detailed information, specifically the concepts present in the aforementioned interface terminology, by applying a blue background. To calibrate the LLMs' summaries for our simplification goal, we chose GPT-4o and used prompt engineering to ensure high-quality prompts and address issues of output inconsistency and prompt sensitivity. We provided both highlighted and unhighlighted versions of each EHR note along with their corresponding prompts to GPT-4o. Each generated summary was manually evaluated to assess its quality using the following evaluation metrics: completeness, correctness, and structural integrity. RESULTS: We used the study sample of 15 discharge notes. On average, summaries from highlighted notes (H-summaries) achieved 96% completeness, 8% higher than the summaries from unhighlighted notes (U-summaries). H-summaries had higher completeness in 13 notes, and U-summaries had higher or equal completeness in 2 notes, resulting in P=.01, which implied statistical significance. Moreover, H-summaries demonstrated better correctness than U-summaries, with fewer instances of erroneous information (2 vs 3 errors, respectively). The number of improper headers was smaller for H-summaries for 11 notes and U-summaries for 4 notes (P=.03; implying statistical significance). Moreover, we identified 8 instances of misplaced information in the U-summaries and only 2 in the H-summaries. We showed that our findings supported the hypothesis that summarizing highlighted discharge notes improves the accuracy of the summaries. CONCLUSIONS: Feeding LLMs with highlighted discharge notes, combined with prompt engineering, results in higher-quality summaries in terms of correctness, completeness, and structural integrity compared to unhighlighted discharge notes.
Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.
Comment cette classification a été obtenuedéplier
Prédiction distillée sur la base complète
Imitation des enseignantsNi prévalence calibrée, ni vérité terrain. Validation humaine à venir. Apprise à partir de 10 348 étiquettes directes de Codex et de 10 348 étiquettes directes de Gemma. Le mode candidate est l'union des têtes enseignantes seuillées; le consensus est leur intersection. Ces sorties portent le statut machine_predicted_unvalidated et ne sont ni des étiquettes humaines ni des étiquettes directes de modèles de pointe.
Scores Codex et Gemma par catégorie
| Catégorie | Codex | Gemma |
|---|---|---|
| Métarecherche | 0,001 | 0,001 |
| Méta-épidémiologie (sens strict) | 0,000 | 0,000 |
| Méta-épidémiologie (sens large) | 0,000 | 0,000 |
| Bibliométrie | 0,000 | 0,001 |
| Études des sciences et des technologies | 0,000 | 0,000 |
| Communication savante | 0,000 | 0,001 |
| Science ouverte | 0,001 | 0,001 |
| Intégrité de la recherche | 0,000 | 0,000 |
| Charge utile insuffisante (le modèle a refusé de juger) | 0,000 | 0,000 |
Scores machine (provisoires)
Les deux têtes enseignantes du modèle étudiant, lues sur ce travail. Un score ordonne la base pour la relecture; il n'affirme jamais une catégorie, et le statut de validation accompagne chaque rangée tel quel.
Scores de référence d'un modèle non mature (critères de maturité non atteints, 7 itérations). Un score ordonne; il n'affirme jamais une catégorie.
score_only:v0-immature-baseline · tel quel depuis la passe de notation : score_only signifie que le nombre peut ordonner les travaux, et qu'aucune étiquette de catégorie n'en découleClassification
machine, non validéePrédiction automatique; un appel candidat d’une seule tête enseignante, pas un consensus.
Le détail, modèle par modèle et score par score, se trouve en fin de page sous « Comment cette classification a été obtenue ».