Less than four years into the large language model era, most if not all who make a living in healthcare are well aware of AI’s growing role in their work. In fact, the breadth and pace of the technology’s advances since late 2022 have been enough to overwhelm even the most clued-in AI aficionado.
Meanwhile “regular” AI hasn’t shifted into neutral to make room for its LLM offspring. And the human brain can be forgiven for clocking slower retrieval times after ingesting deluges of data every day.
The information overload came to mind this week when HealthExec happened upon a brief overview highlighting healthcare-specific AI applications and best practices.
With a hat tip to the authors—members of the staff at advanced IT supplier Databricks—here are excerpts from five of the post’s dozen or so nuggets about AI in healthcare.
1. Electronic health records are the primary data source for most healthcare AI systems.
‘EHR data includes structured fields such as lab results and medication lists alongside unstructured clinical notes, making it a rich but heterogeneous training data source. AI can analyze electronic health records to predict disease risks, identify care gaps and flag patients who may benefit from earlier intervention.’
2. AI systems require high-quality data.
‘It’s needed to ensure accuracy and fairness, which makes data collection standards foundational to any healthcare AI program. Clinical data collection standards should specify acceptable data sources, minimum sample sizes for training data, and documentation requirements for how each dataset was assembled and labeled.’
3. The FDA is planning to monitor AI-equipped medical devices for continuous learning after initial clearance.
‘This represents a departure from the traditional device approval model built around a fixed, unchanging product. A practical compliance checklist for healthcare AI deployments should confirm HIPAA compliance safeguards, applicable AI Act risk classification, FDA clearance status where relevant and documented human oversight procedures before go-live.’
4. Healthcare AI adoption still faces basic challenges like data privacy and algorithmic bias.
‘Addressing both is a prerequisite for responsible deployment rather than an optional add-on. Bias mitigation strategies should include auditing training data for representativeness across demographic groups, testing model performance separately within subpopulations rather than relying on aggregate accuracy alone and establishing a process to retrain or retire models that show a performance gap for any patient group.’
5. AI-driven healthcare is reshaping the skills healthcare professionals need.
‘Interdisciplinary roles are emerging at the intersection of clinical medicine and data science. These include clinical informaticists, health data governance specialists and AI implementation leads responsible for pilot-to-scale processes.’
Commenting on the latter point, the Databricks analysts state: “Healthcare systems investing in these capabilities now are better positioned to evaluate new AI tools critically as healthcare AI innovation accelerates.’