Creation of Scientific Response Documents for Addressing Product Medical Information Inquiries: Mixed Method Approach Using Artificial Intelligence
Notice bibliographique
Résumé
BACKGROUND: Pharmaceutical manufacturers address health care professionals' information needs through scientific response documents (SRDs), offering evidence-based answers to medication and disease state questions. Medical information departments, staffed by medical experts, develop SRDs that provide concise summaries consisting of relevant background information, search strategies, clinical data, and balanced references. With an escalating demand for SRDs and the increasing complexity of therapies, medical information departments are exploring advanced technologies and artificial intelligence (AI) tools like large language models (LLMs) to streamline content development. While AI and LLMs show promise in generating draft responses, a synergistic approach combining an LLM with traditional machine learning classifiers in a series of human-supervised and -curated steps could help address limitations, including hallucinations. This will ensure accuracy, context, traceability, and accountability in the development of the concise clinical data summaries of an SRD. OBJECTIVE: This study aims to quantify the challenges of SRD development and develop a framework exploring the feasibility and value addition of integrating AI capabilities in the process of creating concise summaries for an SRD. METHODS: To measure the challenges in SRD development, a survey was conducted by phactMI, a nonprofit consortium of medical information leaders in the pharmaceutical industry, assessing aspects of SRD creation among its member companies. The survey collected data on the time and tediousness of various activities related to SRD development. Another working group, consisting of medical information professionals and data scientists, used AI to aid SRD authoring, focusing on data extraction and abstraction. They used logistic regression on semantic embedding features to train classification models and transformer-based summarization pipelines to generate concise summaries. RESULTS: Of the 33 companies surveyed, 64% (21/33) opened the survey, and 76% (16/21) of those responded. On average, medical information departments generate 614 new documents and update 1352 documents each year. Respondents considered paraphrasing scientific articles to be the most tedious and time-intensive task. In the project's second phase, sentence classification models showed the ability to accurately distinguish target categories with receiver operating characteristic scores ranging from 0.67 to 0.85 (all P<.001), allowing for accurate data extraction. For data abstraction, the comparison of the bilingual evaluation understudy (BLEU) score and semantic similarity in the paraphrased texts yielded different results among reviewers, with each preferring different trade-offs between these metrics. CONCLUSIONS: This study establishes a framework for integrating LLM and machine learning into SRD development, supported by a pharmaceutical company survey emphasizing the challenges of paraphrasing content. While machine learning models show potential for section identification and content usability assessment in data extraction and abstraction, further optimization and research are essential before full-scale industry implementation. The working group's insights guide an AI-driven content analysis; address limitations; and advance efficient, precise, and responsive frameworks to assist with pharmaceutical SRD development.
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 machine sur la base complète
Imitation des enseignantsNi prévalence calibrée, ni vérité terrain. Validation humaine à venir. Le volet Gemma est une étiquette directe du modèle pour chaque travail de la base, lue sur la notice réduite au titre. Le volet Codex est un classifieur appris des 10 348 étiquettes directes de Codex et calibré sur les taux pondérés de l'échantillon; les champs sans appui suffisant ne portent aucun appel Codex. Le mode candidate est l'union des deux volets; le consensus est leur intersection. Ces sorties portent le statut machine_predicted_unvalidated et ne sont pas des étiquettes humaines.
Scores du classifieur distillé par catégorie (deux têtes)
| Catégorie | Codex | Gemma |
|---|---|---|
| Métarecherche | 0,204 | 0,363 |
| Méta-épidémiologie (sens strict) | 0,003 | 0,003 |
| Méta-épidémiologie (sens large) | 0,002 | 0,004 |
| Bibliométrie | 0,013 | 0,009 |
| Études des sciences et des technologies | 0,004 | 0,003 |
| Communication savante | 0,009 | 0,006 |
| Science ouverte | 0,007 | 0,009 |
| Intégrité de la recherche | 0,003 | 0,003 |
| Charge utile insuffisante (le modèle a refusé de juger) | 0,012 | 0,003 |
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 source (Gemma direct ou Codex distillé), 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 ».