Developed by OpenAI, ChatGPT is a sophisticated large language model (LLM) capable of generating responses that resemble human language when presented with written prompts, making it suitable for various applications. Since its inception, ChatGPT has shown the ability to transform how humans and machines interact, inspiring various applications across multiple domains, including pharmacovigilance. Pharmacovigilance's primary objective is to ensure the safe and efficient utilization of medications while safeguarding public health and patient safety. An integral component of pharmacovigilance is collecting and analyzing safety information related to drugs [1]. ChatGPT's proficiency in handling copious amounts of textual data and its capability to engage in instantaneous conversations with users presents an opportunity to enhance the reporting of adverse drug reactions (ADRs) and improve the accuracy and promptness of pharmacovigilance operations. With its training on a vast 570 GB corpus of diverse online resources, ChatGPT can function as a database for pharmacovigilance lexicon. The significance of ChatGPT lies in its capability to identify ADRs using real-world evidence sourced from nontraditional platforms, such as social media. The US FDA has approximated that a mere 1–10% of all ADRs are reported to the FDA Adverse Event Reporting System (FAERS) [2], whereas conversations about ADRs happen more frequently on social media. Previous research has showcased the capacity of natural language processing (NLP) models in mining ADRs from social media platforms through textual analysis [3]. ChatGPT's ability to utilize its vast general language knowledge gained from training to quickly adapt to new domains with minimal fine-tuning suggests that it would perform better in identifying ADR keywords on platforms such as social media, where informal language is frequently used to describe ADRs [4]. As a preliminary assessment (2 March 2023), we examined ChatGPT's ability to detect drug abuse risk in tweets by comparing its performance against the examples in a published study [4]. When supplying the same set of examples in Table 1 of the study, ChatGPT demonstrated evaluations of drug abuse risk that conform with the results in the table. ChatGPT also demonstrated its ability to provide concise summaries in response to ADR-related inquiries about frequently used medications, with most of the content corroborated by published information. As demonstrated in a test case (2 March 2023), ChatGPT provided a list of adverse effects aligned with a published study for Lasix, a diuretic medication commonly prescribed since 1966 [5]. However, the effectiveness of ChatGPT is significantly influenced by the phrasing of the inquiry. When referring to the drug's brand name, Lasix, ChatGPT can produce a precise ADR inventory. However, when the International Union of Pure and Applied Chemistry (IUPAC) name of the ingredient (4-chloro-[(2-furan-2-ylmethyl)amine]-5-sulfamoylbenzoic acid) was utilized, ChatGPT erroneously associated it with an antibiotic called furazolidone. Nevertheless, a basic internet search by humans using the IUPAC name was able to obtain the intended outcome. Additionally, despite being marketed as a multilingual tool, ChatGPT lacks adequate training data for pharmaceuticals in languages other than English. This deficiency was evident when attempting to retrieve ADR information for the widely used medication Motrin (ibuprofen) in Chinese. In two separate attempts (2 March 2023), ChatGPT was unable to identify Motrin and erroneously linked it with aspirin. The efficacy of ChatGPT in various pharmacovigilance tasks pertaining to less prevalent or newly approved medications that require urgent postmarket monitoring depends on the scientific rigor necessary for each task. Alpelisib, approved for metastatic breast cancer treatment just before ChatGPT's training data cut-off in 2021 [6, 7], was chosen Hanyin Wang and Yanyi Jenny Ding contributed equally to this work.
Paper
Full text
Future of ChatGPT in Pharmacovigilance
Semantic Scholar · Medicine · 2023
Abstract
Developed by OpenAI, ChatGPT is a sophisticated large language model (LLM) capable of generating responses that resemble human language when presented with written prompts, making it suitable for various applications. Since its inception, ChatGPT has shown the ability to transform how humans and machines interact, inspiring various applications across multiple domains, including pharmacovigilance. Pharmacovigilance's primary objective is to ensure the safe and efficient utilization of medications while safeguarding public health and patient safety. An integral component of pharmacovigilance is collecting and analyzing safety information related to drugs [1]. ChatGPT's proficiency in handling copious amounts of textual data and its capability to engage in instantaneous conversations with users presents an opportunity to enhance the reporting of adverse drug reactions (ADRs) and improve the accuracy and promptness of pharmacovigilance operations. With its training on a vast 570 GB corpus of diverse online resources, ChatGPT can function as a database for pharmacovigilance lexicon. The significance of ChatGPT lies in its capability to identify ADRs using real-world evidence sourced from nontraditional platforms, such as social media. The US FDA has approximated that a mere 1–10% of all ADRs are reported to the FDA Adverse Event Reporting System (FAERS) [2], whereas conversations about ADRs happen more frequently on social media. Previous research has showcased the capacity of natural language processing (NLP) models in mining ADRs from social media platforms through textual analysis [3]. ChatGPT's ability to utilize its vast general language knowledge gained from training to quickly adapt to new domains with minimal fine-tuning suggests that it would perform better in identifying ADR keywords on platforms such as social media, where informal language is frequently used to describe ADRs [4]. As a preliminary assessment (2 March 2023), we examined ChatGPT's ability to detect drug abuse risk in tweets by comparing its performance against the examples in a published study [4]. When supplying the same set of examples in Table 1 of the study, ChatGPT demonstrated evaluations of drug abuse risk that conform with the results in the table. ChatGPT also demonstrated its ability to provide concise summaries in response to ADR-related inquiries about frequently used medications, with most of the content corroborated by published information. As demonstrated in a test case (2 March 2023), ChatGPT provided a list of adverse effects aligned with a published study for Lasix, a diuretic medication commonly prescribed since 1966 [5]. However, the effectiveness of ChatGPT is significantly influenced by the phrasing of the inquiry. When referring to the drug's brand name, Lasix, ChatGPT can produce a precise ADR inventory. However, when the International Union of Pure and Applied Chemistry (IUPAC) name of the ingredient (4-chloro-[(2-furan-2-ylmethyl)amine]-5-sulfamoylbenzoic acid) was utilized, ChatGPT erroneously associated it with an antibiotic called furazolidone. Nevertheless, a basic internet search by humans using the IUPAC name was able to obtain the intended outcome. Additionally, despite being marketed as a multilingual tool, ChatGPT lacks adequate training data for pharmaceuticals in languages other than English. This deficiency was evident when attempting to retrieve ADR information for the widely used medication Motrin (ibuprofen) in Chinese. In two separate attempts (2 March 2023), ChatGPT was unable to identify Motrin and erroneously linked it with aspirin. The efficacy of ChatGPT in various pharmacovigilance tasks pertaining to less prevalent or newly approved medications that require urgent postmarket monitoring depends on the scientific rigor necessary for each task. Alpelisib, approved for metastatic breast cancer treatment just before ChatGPT's training data cut-off in 2021 [6, 7], was chosen Hanyin Wang and Yanyi Jenny Ding contributed equally to this work.
References (17)
Scroll for more · 5 remaining