This issue of the North Carolina Medical Journal offers the reader a step-by-step journey into artificial intelligence as a tool to improve the outputs and outcomes of clinical practice. While the predictions and promise are grand, limitations, cautions, and guardrails are also described.
Consider electronic health records. Machine learning can rapidly scan, search, and pull information from the rich free text in admission notes, social histories, problem lists, progress notes, and discharge summaries across different users and platforms to identify common threads. Machine learning becomes artificial intelligence when the search identifies the “story” of disease progression detailed in progress notes linked to diagnostic tests, vital signs, medications, and more, to reveal patterns that affect prognoses and outcomes.
AI can help create intentional case studies utilizing data from hundreds to thousands of participants, seeking patterns that inform clinical care ranging from, for example, antibiotic stewardship to pharmacologic wisdom to early discernment of markers of shock and sepsis. AI, too, can search the broad clinical literature to identify which treatment modalities might best be applied to the population with these signs and symptoms and diagnostic findings. It can analyze medical images, revolutionize clinical trials, and improve claims management. But it is here that AI needs empathy, insight, and humility. If the searchable data are skewed by gender, race, ethnicity, or socioeconomic or insurance status, to name a few variables, conclusions and recommendations might be biased and not applicable to an under-represented population in the data set.
As authors in this issue explain, we must approach this multitude of opportunities and this exciting frontier with caution. Their overall message: “plan, don’t panic.”