A 30-second retinal photo, a few spoken sentences, or 15 seconds of a heartbeat can now feed an AI model that flags hidden disease, from diabetic retinopathy to heart failure to early Parkinson’s. These systems are extraordinary at one thing: sifting enormous volumes of subtle data for patterns no human eye can reliably catch. But across every FDA-authorized example, the model produces a screen, a score, or a referral, and a clinician makes the diagnosis and decides what happens next. This is decision support, not decision replacement.
Key takeaways
- Eyes as a window: A Google/Verily model predicted cardiovascular risk factors and 5-year heart-attack/stroke risk from retinal photographs alone, trained on 284,335 patients (Nature Biomedical Engineering, 2018).
- First autonomous diagnostic: IDx-DR (now LumineticsCore) became the first FDA-authorized autonomous AI diagnostic in any field in 2018, detecting diabetic retinopathy at 87% sensitivity and 90% specificity in primary care.
- Voice is a biomarker: Speech changes can precede Parkinson’s motor symptoms by up to a decade; research-grade voice models report high accuracy, but no voice tool is yet FDA-cleared for autonomous diagnosis.
- Heartbeat in 15 seconds: Eko Health’s AI stethoscope, FDA-cleared in 2024, flags low ejection fraction during a routine exam, trained on 100,000+ ECG/echo pairs.
- A growing but contested market: AI in diagnostics is estimated at roughly US$2 billion in 2025, with 2030 forecasts varying widely (one common figure: $5.44B).
Source: Nature Biomedical Engineering, 2018
Source: npj Digital Medicine, 2018
Source: Eko Health / FDA, 2024
Source: Grand View Research, 2025
Retinal scans that read your heart
Researchers at Google and Verily trained deep-learning models on retinal fundus photographs and health data from 284,335 patients, then showed the AI could estimate cardiovascular risk factors directly from the eye, including age, blood pressure, smoking status, and a patient’s 5-year risk of a major adverse cardiac event such as heart attack or stroke. The model distinguished smokers from non-smokers about 71% of the time and predicted MACE with an area-under-curve around 0.70, comparable to established but more invasive risk calculators. Crucially, the authors framed it as a non-invasive screening signal to surface high-risk patients for a clinician’s review, not as a standalone verdict. Source: Nature Biomedical Engineering, 2018.
IDx-DR: the first autonomous AI diagnostic
In April 2018, the FDA granted De Novo authorization to IDx-DR (now marketed as LumineticsCore by Digital Diagnostics), the first autonomous AI diagnostic system cleared in any field of medicine. In its pivotal trial across 10 primary-care sites and 900 patients with diabetes, the system detected more-than-mild diabetic retinopathy at 87% sensitivity, 90% specificity, and a 96% imageability rate, all from retinal images interpreted without a specialist on site. Even in this “autonomous” case, a positive result is a referral instruction: it tells the primary-care team to send the patient to an eye specialist for diagnosis and treatment. Source: npj Digital Medicine, 2018.
Your voice as a biomarker
Parkinson’s disease produces hypokinetic dysarthria, subtle changes in vocal loudness, pitch monotonicity, and articulation, that can appear up to a decade before overt motor symptoms. Machine-learning models built on acoustic features (such as pitch perturbation and recurrence-period density) have reported high research-grade accuracy, with some ensemble and deep-learning studies citing AUC values above 0.98 on benchmark voice datasets. These figures come from controlled research samples and should be read cautiously: no voice-based tool is yet FDA-cleared as an autonomous diagnostic, and current efforts position voice as a low-cost screening prompt for clinical follow-up. Source: Scientific Reports, 2025.
A heartbeat-and-sepsis double in 2024
Two 2024 FDA actions show how fast clearance-grade tools are arriving. In April 2024 the FDA cleared Eko Health’s Low EF AI, developed with the Mayo Clinic, which analyzes ECG and heart-sound data captured by a digital stethoscope in about 15 seconds to flag low ejection fraction, a key heart-failure sign; it was trained on more than 100,000 paired ECG/echocardiogram exams and reported roughly 74.7% sensitivity and 77.5% specificity. Days earlier, the FDA granted Prenosis De Novo authorization for the Sepsis ImmunoScore, the first AI diagnostic for sepsis, which combines 22 parameters into a risk score across four categories. Notably, the company stresses it is “not an alert system” but a decision aid that a clinician interprets alongside the full clinical picture. Source: Eko Health / FDA, 2024.
Market and impact: real growth, fuzzy numbers
The AI-in-diagnostics market is growing quickly, but published estimates diverge enough to warrant caution. Grand View Research pegs the market at roughly US$1.97 billion in 2025, rising to about $5.44 billion by 2030 at a CAGR near 22.5%, while other analysts report CAGRs from 23% to nearly 25% and different base years, and broader “AI in medical diagnostics” definitions push totals far higher. The takeaway is directional rather than precise: strong double-digit growth driven by clinician shortages and rising chronic-disease burden, with the exact dollar figure depending heavily on how each firm scopes the category. Source: Grand View Research, 2025.
The through-line: decision support, not replacement
Across retinal imaging, voice, and the heart, the pattern is consistent: AI excels at processing data volume and detecting faint signals, then hands a clinician a flag, score, or referral. Even IDx-DR, the most “autonomous” example, outputs a recommendation to refer, not a treatment plan. Voice models stay in screening territory; sepsis and heart-failure tools are explicitly framed as aids a doctor weighs against the whole patient. The clinical and regulatory consensus is the same as the brand thesis: the algorithm widens who gets screened and how early, but a human still makes the final call.
Methodology & sources
- Retinal cardiovascular-risk prediction from 284,335 patients — Nature Biomedical Engineering (2018)
- IDx-DR pivotal trial: 87% sensitivity, 90% specificity, 96% imageability — npj Digital Medicine (2018)
- FDA De Novo authorization details for IDx-DR — FDA DEN180001 (2018)
- Voice biomarkers for early Parkinson’s detection — Scientific Reports (2025)
- Eko Health Low EF AI FDA clearance and performance — Eko Health / FDA (2024)
- Sepsis ImmunoScore, first FDA-authorized AI sepsis diagnostic — Prenosis (2024)
- AI-in-diagnostics market sizing and CAGR — Grand View Research (2025)
Frequently asked questions
Can AI diagnose disease from a photo of my eye?
AI can screen for disease from a retinal photo, and one system (IDx-DR/LumineticsCore) is FDA-authorized to autonomously detect diabetic retinopathy. Research models have also estimated cardiovascular risk factors from retinal images. In practice these tools flag or refer patients, and a clinician confirms the diagnosis and decides on treatment.
Can your voice reveal a medical condition?
Yes, voice can act as a biomarker, because conditions like Parkinson’s disease subtly alter speech years before motor symptoms appear. Research-grade AI models report high accuracy detecting these vocal changes, but no voice tool is yet FDA-cleared for autonomous diagnosis, so today it functions as a screening prompt for clinical follow-up.
Does AI replace the doctor in these diagnoses?
No, AI does not replace the doctor; it provides decision support. The AI handles data volume and pattern detection to produce a screen, score, or referral, and a human clinician makes the final diagnostic and treatment decision. Even FDA-authorized autonomous tools output a recommendation to refer rather than a complete care plan.
Part of our Real-World AI Use Cases series — how AI supports high-stakes decisions across surprising domains.