There is great excitement that medical artificial intelligence (AI) based on machine learning (ML) can be used to improve decision making at the patient level in a variety of healthcare settings. However, the quantification and communication of uncertainty for individual predictions is often neglected even though uncertainty estimates could lead to more principled decision-making and enable machine learning models to automatically or semi-automatically abstain on samples for which there is high uncertainty. In this article, we provide an overview of different approaches to uncertainty quantification and abstention for machine learning and highlight how these techniques could improve the safety and reliability of current ML systems being used in healthcare settings. Effective quantification and communication of uncertainty could help to engender trust with healthcare workers, while providing safeguards against known failure modes of current machine learning approaches. As machine learning becomes further integrated into healthcare environments, the ability to say “I’m not sure” or “I don’t know” when uncertain is a necessary capability to enable safe clinical deployment.

Word embeddings are a popular approach to unsupervised learning of word relationships that are widely used in natural language processing. In this article, we present a new set of embeddings for medical concepts learned using an extremely large collection of multimodal medical data. Leaning on recent theoretical insights, we demonstrate how an insurance claims database of 60 million members, a collection of 20 million clinical notes, and 1.7 million full text biomedical journal articles can be combined to embed concepts into a common space, resulting in the largest ever set of embeddings for 108,477 medical concepts. To evaluate our approach, we present a new benchmark methodology based on statistical power specifically designed to test embeddings of medical concepts. Our approach, called cui2vec, attains state-of-the-art performance relative to previous methods in most instances. Finally, we provide a downloadable set of pre-trained embeddings for other researchers to use, as well as an online tool for interactive exploration of the cui2vec embeddings.

Importance: Artificial intelligence (AI) applications in health care have been effective in many areas of medicine, but they are often trained for a single task using labeled data, making deployment and generalizability challenging. Whether a general-purpose AI language model can perform diagnosis and triage is unknown. Objective: Compare the general-purpose Generative Pre-trained Transformer 3 (GPT-3) AI model's diagnostic and triage performance to attending physicians and lay adults who use the Internet. Design: We compared the accuracy of GPT-3's diagnostic and triage ability for 48 validated case vignettes of both common (e.g., viral illness) and severe (e.g., heart attack) conditions to lay people and practicing physicians. Finally, we examined how well calibrated GPT-3's confidence was for diagnosis and triage. Setting and Participants: The GPT-3 model, a nationally representative sample of lay people, and practicing physicians. Exposure: Validated case vignettes (<60 words; <6th grade reading level). Main Outcomes and Measures: Correct diagnosis, correct triage. Results: Among all cases, GPT-3 replied with the correct diagnosis in its top 3 for 88% (95% CI, 75% to 94%) of cases, compared to 54% (95% CI, 53% to 55%) for lay individuals (p<0.001) and 96% (95% CI, 94% to 97%) for physicians (p=0.0354). GPT-3 triaged (71% correct; 95% CI, 57% to 82%) similarly to lay individuals (74%; 95% CI, 73% to 75%; p=0.73); both were significantly worse than physicians (91%; 95% CI, 89% to 93%; p<0.001). As measured by the Brier score, GPT-3 confidence in its top prediction was reasonably well-calibrated for diagnosis (Brier score = 0.18) and triage (Brier score = 0.22). Conclusions and Relevance: A general-purpose AI language model without any content-specific training could perform diagnosis at levels close to, but below physicians and better than lay individuals. The model was performed less well on triage, where its performance was closer to that of lay individuals.

No abstract