AI in Primary Care

Primary care AI falls into four broad jobs: screening, chronic disease tracking, documentation and triage. Screening is where AI first earned regulatory trust. In 2018 the US FDA authorized IDx-DR (now called LumineticsCore) as the first autonomous AI diagnostic system: it reads retinal photos taken in a primary care office and decides whether a specialist referral is needed, without a clinician interpreting the image. That design matters because many people with diabetes never attend separate eye appointments. Researchers at Mayo Clinic also tested an AI that flags likely weak heart pumping (low ejection fraction) from an ordinary 12-lead ECG; in a pragmatic trial in primary care practices, giving clinicians the alert increased new diagnoses of the condition. Chronic disease tracking uses the electronic health record. Models sort patients by risk of hospital admission, uncontrolled diabetes or missed care so that limited nurse time goes to the right people. These are usually population tools, not bedside diagnoses. Documentation is the fastest-growing use. Ambient scribes such as Microsoft's DAX Copilot and Abridge record the conversation and produce a draft note using speech recognition and a large language model. Large health systems have rolled them out widely because note-writing after hours is a major cause of burnout. The clinician remains responsible for every word, and drafts can contain omissions or invented details. Triage covers symptom checkers and message sorting. This area has a mixed record: consumer symptom checkers have shown inconsistent accuracy, and the UK company Babylon Health, which promoted an AI triage chatbot, collapsed in 2023. A common misconception is that these tools replace the GP. In practice most are narrow assistants whose value depends on whether a human acts on their output and whether the practice changes its workflow around them.