Diagnostic Accuracy of Differential-Diagnosis Lists Generated by Generative Pretrained Transformer 3 Chatbot for Clinical Vignettes with Common Chief Complaints: A Pilot Study

There is no author summary for this article yet. Authors can add summaries to their articles on ScienceOpen to make them more accessible to a non-specialist audience.

Abstract

The diagnostic accuracy of differential diagnoses generated by artificial intelligence (AI) chatbots, including the generative pretrained transformer 3 (GPT-3) chatbot (ChatGPT-3) is unknown. This study evaluated the accuracy of differential-diagnosis lists generated by ChatGPT-3 for clinical vignettes with common chief complaints. General internal medicine physicians created clinical cases, correct diagnoses, and five differential diagnoses for ten common chief complaints. The rate of correct diagnosis by ChatGPT-3 within the ten differential-diagnosis lists was 28/30 (93.3%). The rate of correct diagnosis by physicians was still superior to that by ChatGPT-3 within the five differential-diagnosis lists (98.3% vs. 83.3%, p = 0.03). The rate of correct diagnosis by physicians was also superior to that by ChatGPT-3 in the top diagnosis (53.3% vs. 93.3%, p < 0.001). The rate of consistent differential diagnoses among physicians within the ten differential-diagnosis lists generated by ChatGPT-3 was 62/88 (70.5%). In summary, this study demonstrates the high diagnostic accuracy of differential-diagnosis lists generated by ChatGPT-3 for clinical cases with common chief complaints. This suggests that AI chatbots such as ChatGPT-3 can generate a well-differentiated diagnosis list for common chief complaints. However, the order of these lists can be improved in the future.

Related collections

Most cited references 33

Record: found
Abstract: found
Article: not found

Language Models are Few-Shot Learners

Tom B. Brown, Benjamin Mann, Nick Ryder … (2020)

Recent work has demonstrated substantial gains on many NLP tasks and benchmarks by pre-training on a large corpus of text followed by fine-tuning on a specific task. While typically task-agnostic in architecture, this method still requires task-specific fine-tuning datasets of thousands or tens of thousands of examples. By contrast, humans can generally perform a new language task from only a few examples or from simple instructions - something which current NLP systems still largely struggle to do. Here we show that scaling up language models greatly improves task-agnostic, few-shot performance, sometimes even reaching competitiveness with prior state-of-the-art fine-tuning approaches. Specifically, we train GPT-3, an autoregressive language model with 175 billion parameters, 10x more than any previous non-sparse language model, and test its performance in the few-shot setting. For all tasks, GPT-3 is applied without any gradient updates or fine-tuning, with tasks and few-shot demonstrations specified purely via text interaction with the model. GPT-3 achieves strong performance on many NLP datasets, including translation, question-answering, and cloze tasks, as well as several tasks that require on-the-fly reasoning or domain adaptation, such as unscrambling words, using a novel word in a sentence, or performing 3-digit arithmetic. At the same time, we also identify some datasets where GPT-3's few-shot learning still struggles, as well as some datasets where GPT-3 faces methodological issues related to training on large web corpora. Finally, we find that GPT-3 can generate samples of news articles which human evaluators have difficulty distinguishing from articles written by humans. We discuss broader societal impacts of this finding and of GPT-3 in general. 40+32 pages

0 comments Cited 644 times – based on 0 reviews      Review now

Bookmark

Record: found
Abstract: found
Article: not found

Semantics derived automatically from language corpora contain human-like biases

Aylin Caliskan, Joanna Bryson, Arvind Narayanan (2017)

Machine learning is a means to derive artificial intelligence by discovering patterns in existing data. Here, we show that applying machine learning to ordinary human language results in human-like semantic biases. We replicated a spectrum of known biases, as measured by the Implicit Association Test, using a widely used, purely statistical machine-learning model trained on a standard corpus of text from the World Wide Web. Our results indicate that text corpora contain recoverable and accurate imprints of our historic biases, whether morally neutral as toward insects or flowers, problematic as toward race or gender, or even simply veridical, reflecting the status quo distribution of gender with respect to careers or first names. Our methods hold promise for identifying and addressing sources of bias in culture, including technology.

0 comments Cited 259 times – based on 0 reviews      Review now

Bookmark

Record: found
Abstract: found
Article: found

Is Open Access

Evaluation of symptom checkers for self diagnosis and triage: audit study

Hannah Semigran, Jeffrey Linder, Courtney Gidengil … (2015)

Objective To determine the diagnostic and triage accuracy of online symptom checkers (tools that use computer algorithms to help patients with self diagnosis or self triage). Design Audit study. Setting Publicly available, free symptom checkers. Participants 23 symptom checkers that were in English and provided advice across a range of conditions. 45 standardized patient vignettes were compiled and equally divided into three categories of triage urgency: emergent care required (for example, pulmonary embolism), non-emergent care reasonable (for example, otitis media), and self care reasonable (for example, viral upper respiratory tract infection). Main outcome measures For symptom checkers that provided a diagnosis, our main outcomes were whether the symptom checker listed the correct diagnosis first or within the first 20 potential diagnoses (n=770 standardized patient evaluations). For symptom checkers that provided a triage recommendation, our main outcomes were whether the symptom checker correctly recommended emergent care, non-emergent care, or self care (n=532 standardized patient evaluations). Results The 23 symptom checkers provided the correct diagnosis first in 34% (95% confidence interval 31% to 37%) of standardized patient evaluations, listed the correct diagnosis within the top 20 diagnoses given in 58% (55% to 62%) of standardized patient evaluations, and provided the appropriate triage advice in 57% (52% to 61%) of standardized patient evaluations. Triage performance varied by urgency of condition, with appropriate triage advice provided in 80% (95% confidence interval 75% to 86%) of emergent cases, 55% (47% to 63%) of non-emergent cases, and 33% (26% to 40%) of self care cases (P<0.001). Performance on appropriate triage advice across the 23 individual symptom checkers ranged from 33% (95% confidence interval 19% to 48%) to 78% (64% to 91%) of standardized patient evaluations. Conclusions Symptom checkers had deficits in both triage and diagnosis. Triage advice from symptom checkers is generally risk averse, encouraging users to seek care for conditions where self care is reasonable.