Dr Neil Panchal explains that while the current healthcare landscape offers significant opportunities for AI in chronic disease care, achieving better patient outcomes relies on the effective governance of the underlying technology
Modern medicine has an inversion problem. For most of the twentieth century, clinicians made decisions with too little information. Today they drown in it. A single patient with three chronic conditions may accumulate thousands of data points a year across electronic health records, imaging archives, laboratory networks, pharmacy claims, and increasingly, consumer wearables. Longitudinal history – the connective tissue of chronic disease care – is scattered across systems that were never designed to speak to each other. The consequences are predictable and expensive: redundant testing, missed abnormal results, treatment inertia, and clinicians burning cognitive energy on data reconciliation rather than judgment. The fundamental clinical problem is no longer data scarcity. It is synthesis.
The federal posture has begun to catch up. On July 30, 2025, the Centers for Medicare and Medicaid Services announced the Health Technology Ecosystem initiative, establishing voluntary interoperability commitments across payers, health systems, and technology platforms. In December 2025, the CMS Innovation Center followed with the Advancing Chronic Care with Effective, Scalable Solutions (ACCESS) Model – a ten-year, outcome-aligned payment framework for technology-supported chronic disease management, whose first performance period opened on July 1, 2026. This is meaningful policy signaling. Whether it translates into better outcomes will depend on how the underlying technology is governed
The transformative potential of AI in healthcare
Early detection is where the case for AI is strongest and most concrete. A twelve-lead electrocardiogram is inexpensive and ubiquitous, but until recently was interpreted only for abnormalities discernible to the human eye or encoded in rule-based algorithms. When researchers trained a convolutional neural network on paired ECG-echocardiogram data, the model identified asymptomatic left ventricular dysfunction from ECG waveform alone. (1) The pragmatic EAGLE randomized trial then showed that deploying this algorithm in primary care increased detection of low ejection fraction in real-world workflows. (2) A parallel story is unfolding: deep learning models have been trained to predict structural progression before radiographic thresholds are crossed, (3) opening a window for structured intervention that current standard-of-care misses.
Risk stratification is the second lane. The American Heart Association’s PREVENT equations, released in 2023, expanded traditional cardiovascular risk modeling to include kidney function and social determinants – a substantial improvement, but still a static snapshot of a dynamic biological trajectory. AI-augmented stratification does something different: it integrates polygenic risk scores, coronary artery calcium quantification, longitudinal lipid slopes, and clinically validated calculators simultaneously. A patient with a normal ApoB today but a decade of rising values carries a risk signature no single-visit calculator captures. This is where large-scale pattern recognition, applied to individual longitudinal data, offers physicians something they could not previously operationalize.
Personalization follows. Continuous glucose monitors, activity trackers, sleep sensors, and heart rate variability devices generate the kind of behavioral and physiological signal that clinic-based measurements only glimpse. Layered onto conventional biomarkers – hemoglobin A1c, ApoB, inflammatory markers – these data streams transform treatment planning from episodic to iterative. A diabetic patient with poor glycemic control after two consecutive nights of fragmented sleep is offered a non-pharmacologic intervention before medication is escalated. Hybrid closed-loop insulin systems already prove the architecture at pharmacological scale; the generalization to broader chronic disease management is straightforward.
AI functions best as the tireless resident that surfaces patterns; the physician remains the attending, holding final clinical authority.
Long-term monitoring is where the most durable value likely sits. Chronic disease is a longitudinal problem, and clinicians simply cannot manually track thousands of continuously updating data streams. Agentic AI systems can surface meaningful changes – a rising resting heart rate over ten days, a new pattern of nocturnal hypoglycemia, a lapse in medication refill – and queue them for physician review. Curated, safety-bounded patient-facing conversational tools (which CMS explicitly named in its 2025 announcement) can handle adherence checks, symptom logging, and medication reconciliation, escalating to clinicians only when human judgment is required. The result is not less physician involvement. It is better-directed physician involvement.
Navigating the integration of AI in chronic disease care
Two honest caveats belong in any policy conversation. First, retrospective algorithmic performance often does not survive prospective deployment; bias embedded in training data compounds existing disparities if not audited. Second, alerts without governance produce fatigue, not improvement. The lesson of the past two decades of clinical decision support is that more signal without curation degrades attention.
The strategic question is not whether AI enters chronic disease care. It is how well its authority is bounded. AI functions best as the tireless resident that surfaces patterns; the physician remains the attending, holding final clinical authority. Policy that funds interoperability, mandates prospective evidence, and enforces clinician-in-the-loop design will decide whether this generation of tools shortens diagnostic delay – or repeats the disappointments of the last one.
References
- Attia ZI et al. Screening for cardiac contractile dysfunction using an artificial intelligence-enabled electrocardiogram. Nat Med. 2019 Jan;25(1):70-74. doi: 10.1038/s41591-018-0240-2. Epub 2019 Jan 7. PMID: 30617318.
- Yao X et al. Artificial intelligence-enabled electrocardiograms for identification of patients with low ejection fraction: a pragmatic, randomized clinical trial. Nat Med. 2021 May;27(5):815-819. doi: 10.1038/s41591-021-01335-4. Epub 2021 May 6. PMID: 33958795.
- Joseph GB et al. Machine learning models for clinical and structural knee osteoarthritis prediction: Recent advancements and future directions. Osteoarthr Cartil Open. 2025 Jul 24;7(3):100654. doi: 10.1016/j.ocarto.2025.100654. PMID: 40799630; PMCID: PMC12341514.
Leave a Reply