Explore the research

Medicine

Does AI improve diagnosis, reduce harm and ultimately make patients healthier?

21%

field summary estimate

6 outcomes · Assessed November 7, 2026 · Low confidence

The field estimate averages its outcome scores. It is a reading aid, not a measured fraction of the field that AI manages.

How the estimate built up.

Each band is one outcome's share of the field estimate, so the top line is the estimate itself. Every field starts at 0% on December 31, 2022. Dates before the latest are retrospective estimates. Point at a band to see that outcome's score on each date.

Patient benefit15%
Harm and unnecessary care20%
Equity and transfer15%
Diagnostic performance35%
Cost effectiveness15%
Care delivery25%

View exact historical values and reconstruction notes

What we measure

Each percentage is a judgment about progress toward independent advancement. Open the evidence to inspect the basis for it.

Care delivery

25%estimated progress

What we track

Time to appropriate treatment, waiting time, completed visits, and clinician minutes per patient. Include time spent correcting AI output.

Why this estimate?

Reduced reading workload demonstrates useful workflow automation, but end-to-end treatment timeliness and human labor across the care pathway remain unmeasured.

Advancements

MASAI trial: AI cuts mammogram reading workload by 44%August 2023. The 2023 MASAI interim analysis reported a 44.3% reduction in screen-reading workload.
ACCESS trial: point-of-care AI eye exam completed by 100% vs 22%January 11, 2024. Randomized trial of 164 youth with diabetes: 100% of those offered an autonomous AI diabetic eye exam at the clinic completed an exam within 6 months, versus 22% of those given a standard referral (p < 0.001).
Danish screening programme: AI cuts mammogram reading workload 33.5%June 2024. Retrospective before-and-after study in the Capital Region of Denmark (60,751 women before AI, 58,246 with AI): radiologist reading workload fell 33.5% (38,977 of 116,492 reads) after AI was introduced.
UCLA trial: Nabla AI scribe cuts physician time-in-note by 9.5%November 26, 2025. Three-group pragmatic randomized trial of 238 outpatient physicians in 14 specialties: time-in-note fell 9.5% versus usual care with Nabla (95% CI -17.2% to -1.8%); the change with DAX Copilot (-1.7%) was not significant.
Pragmatic trial: ambient AI cuts note time by 0.36 hours per dayNovember 26, 2025. A 24-week stepped-wedge, individually randomized trial of 66 practitioners in ambulatory clinics in two states: time on notes fell 0.36 hours per day (95% CI -0.55 to -0.17) and work exhaustion/disengagement fell 0.44 points on a five-point scale.

Read the dated findings and sources

Status: Measured workload proxy. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

August 2023: The 2023 MASAI interim analysis reported a 44.3% reduction in screen-reading workload. MASAI interim randomized trial analysis.

Interpretation and tracking value

Reading workload is not clinician minutes per patient or time to appropriate treatment. The gain must be evaluated alongside recall and diagnostic work downstream.

Next comparable measurement

Measure actual staff minutes, correction/review time, waits and treatment delays for the full pathway, including additional recalls. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Cost effectiveness

15%estimated progress

What we track

Full cost per appropriately managed patient and, where valid, cost per quality-adjusted life-year gained. State assumptions.

Why this estimate?

Economic modeling makes the outcome assessable, but the cited setting did not meet its cost-effectiveness threshold. Autonomous optimization of real patient benefit per dollar is unproven.

Advancements

Urban China study: AI retinopathy screening not cost-effectiveFebruary 23, 2023. Cost-effectiveness analysis with real-world data from community diabetic retinopathy telescreening: the AI-assisted model cost US $15,216.96 per QALY versus manual grading and was judged not cost-effective, though it could become dominant if referral compliance rose 7.5%.
EAGLE trial analysis: AI-ECG screening costs $27,858 per QALYOctober 26, 2024. Secondary cost-effectiveness analysis of the EAGLE randomized trial of AI-enabled ECG screening for low ejection fraction: ICER of $27,858 per QALY versus usual care, and $1,651 per QALY in the outpatient setting.
UK model: AI replacing mammogram readers saves £159.55 per womanMay 2, 2026. Discrete-event simulation of the NHS breast screening pathway using prospective trial evidence: an AI-only reading strategy gained 0.00009 QALYs and cut lifetime costs by £159.55 per woman invited; replacing one reader saved £31.07; adding AI as a third reader raised costs by £72.79.
Study finds $40 AI breast screening is not cost-effective2026 publication. The modeling study concluded that AI screening priced at $40 per examination was not cost-effective at a $100,000-per-QALY willingness-to-pay threshold.
China model: community AI screening for ROP is cost-savingOctober 2, 2026. Decision-tree model of 100,000 preterm infants: community-based AI-assisted screening for retinopathy of prematurity dominated bedside specialist screening, with ICERs of USD -7,307 per QALY (urban) and -3,688 per QALY (rural).

Read the dated findings and sources

Status: Model-dependent economic result. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

2026 publication: The modeling study concluded that AI screening priced at $40 per examination was not cost-effective at a $100,000-per-QALY willingness-to-pay threshold. Breast screening cost-effectiveness modeling study.

Interpretation and tracking value

This is a modeled setting, not an observed universal price threshold; benefits, downstream costs, utility assumptions and local practice can change the conclusion.

Next comparable measurement

Record currency/year, AI price, full pathway costs, incremental QALYs and sensitivity analyses; validate model inputs against prospective local data. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Diagnostic performance

35%estimated progress

What we track

Sensitivity, specificity, positive predictive value, calibration, and missed diagnoses against an independently determined reference standard. Report prevalence and follow-up.

Why this estimate?

Randomized screening evidence supports reliable bounded diagnostic assistance. General clinical transfer, human-free adjudication and self-directed medical advancement remain unestablished.

Advancements

ScreenTrustCAD: one radiologist plus AI non-inferior to two readersSeptember 8, 2023. Prospective paired-reader study of 55,581 women in Sweden: one radiologist plus AI detected 261 cancers versus 250 for standard double reading (relative proportion 1.04, 95% CI 1.00-1.09); AI alone detected 246 (0.98, 95% CI 0.93-1.04), also non-inferior.
PANDA detects pancreatic lesions on non-contrast CT, 92.9% sensitivityNovember 20, 2023. Deep learning model validated in 6,239 patients at 10 centres (AUC 0.986-0.996) and in 20,530 consecutive real-world patients, with 92.9% sensitivity and 99.9% specificity for lesion detection; it exceeded mean radiologist sensitivity by 34.1%.
Trial: GPT-4 access does not improve physicians' diagnostic reasoningOctober 2024. Randomized trial of 50 physicians: median diagnostic reasoning score was 76% with GPT-4 access versus 74% with conventional resources (difference 2 points, 95% CI -4 to 8, P = .60). GPT-4 alone scored 16 points higher than the conventional-resources group (95% CI 2 to 30).
PRAIM: AI-supported screening finds 6.7 vs 5.7 cancers per 1,000January 7, 2025. Real-world implementation study of over 460,000 women at 12 German screening sites: AI-supported double reading detected 6.7 breast cancers per 1,000 women versus 5.7 per 1,000 with standard double reading, with recalls of 37.4 versus 38.3 per 1,000.
AMIE outperforms primary care physicians on 30 of 32 axesApril 9, 2025. In a randomized, blinded study of text consultations with patient-actors across 159 case scenarios, Google's AMIE was rated better than 20 primary care physicians on 30 of 32 axes by specialist physicians and 25 of 26 axes by patient-actors.
Microsoft MAI-DxO reaches 85.5% on 304 NEJM case challengesJune 27, 2025. On 304 NEJM clinicopathological conference cases recast as stepwise diagnostic encounters, MAI-DxO paired with OpenAI o3 reached 80% accuracy (85.5% in its maximum-accuracy configuration) versus about 20% for generalist physicians. This is a benchmark preprint, not a clinical trial.
Penda Health: AI Consult cuts diagnostic errors 16% in 39,849 visitsJuly 22, 2025. Real-world study at 15 primary care clinics in Nairobi: clinicians with the LLM-based AI Consult tool made 16% fewer diagnostic errors and 13% fewer treatment errors (relative) across 39,849 visits, as rated by independent physicians. Preprint with OpenAI involvement.
MASAI: AI-supported screening raises sensitivity to 80.5%January 31, 2026. MASAI reports sensitivity 80.5% (95% CI 76.4–84.2) with AI-supported screening versus 73.8% (68.9–78.3) with standard reading; specificity was 98.5% in both.

Read the dated findings and sources

Status: Measured randomized screening performance. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

January 31, 2026: MASAI reports sensitivity 80.5% (95% CI 76.4–84.2) with AI-supported screening versus 73.8% (68.9–78.3) with standard reading; specificity was 98.5% in both. MASAI randomized screening trial follow-up.

Interpretation and tracking value

This applies to the trial's breast-screening population and workflow, not all diagnostic tasks. Predictive value and calibration require prevalence-specific assessment.

Next comparable measurement

Preserve screening interval and reference standard; report sensitivity, specificity, PPV and calibration with complete follow-up and confidence intervals. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Equity and transfer

15%estimated progress

What we track

Outcome and error differences across relevant demographic groups, institutions, devices, and disease severities.

Why this estimate?

Age and density subgroup evidence supports limited transportability. Broader demographic and institutional outcome equality cannot be inferred from it.

Advancements

Four commercial LLMs all repeat debunked race-based medicineOctober 20, 2023. Four commercially available large language models were each asked nine questions five times (45 responses per model); all four produced examples perpetuating race-based medicine, and answers were not consistent across repeats.
Study: fairness fixes in imaging AI do not transfer to new sitesJune 28, 2024. Across radiology, dermatology and ophthalmology models and six chest X-ray datasets, imaging AI was found to use demographic shortcuts; correcting them closed fairness gaps locally but not in new test settings, where models encoding less demographic information were often fairest.
CONCERN warning system: mortality benefit holds across patient groupsJune 10, 2025. Secondary analysis of the cluster-randomized CONCERN trial at four hospitals: tool usage did not differ by patient demographics; mortality benefit was consistent across most groups (overall adjusted HR 0.644) and larger for patients whose primary language was not English (adjusted HR 0.419, 95% CI 0.287-0.610).
MASAI: sensitivity gains hold across age and breast-density groupsJanuary 31, 2026. MASAI reports sensitivity improvement consistent across age and breast-density groups.

Read the dated findings and sources

Status: Subgroup evidence; wider transportability unknown. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

January 31, 2026: MASAI reports sensitivity improvement consistent across age and breast-density groups. MASAI randomized screening trial follow-up.

Interpretation and tracking value

Within-trial subgroup consistency is not evidence of equal outcomes across ethnicities, countries, scanners or clinical systems; small subgroups may conceal uncertainty.

Next comparable measurement

Validate in additional institutions and devices; report absolute subgroup errors and patient outcomes with uncertainty, including underrepresented populations. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Harm and unnecessary care

20%estimated progress

What we track

Adverse events, false-positive procedures, missed urgent cases, and inappropriate treatments per eligible patient.

Why this estimate?

A screening safety endpoint supports bounded use, while overdiagnosis, downstream procedures and other harms still require comprehensive human-led evaluation.

Advancements

Meta-analysis: colonoscopy AI raises removal of harmless polypsAugust 29, 2023. Meta-analysis of 21 randomized trials (18,232 patients): computer-aided detection raised adenoma detection from 35.9% to 44.0% but also raised non-neoplastic polyp removals from 0.34 to 0.52 per colonoscopy, without increasing detection of advanced adenomas.
Vignette trial: biased AI lowers clinician accuracy by 11.3 pointsDecember 2023. Randomized vignette study of 457 clinicians diagnosing acute respiratory failure: baseline accuracy was 73.0%; a standard AI model raised it 2.9 points, while a systematically biased model lowered it 11.3 points, and image-based explanations did not offset the harm.
Danish screening: false positives fall from 2.39% to 1.63% with AIJune 2024. After AI was introduced into mammography screening in the Capital Region of Denmark, the recall rate fell from 3.09% to 2.46% and the false-positive rate from 2.39% to 1.63%, while cancer detection rose from 0.70% to 0.82% and positive predictive value from 22.6% to 33.6%.
Polish study: endoscopists' unaided detection drops after AI useAugust 12, 2025. Observational study at four Polish centres: in colonoscopies done without AI, the adenoma detection rate fell from 28.4% (226 of 795) before endoscopists were routinely exposed to AI to 22.4% (145 of 648) afterwards, an absolute difference of -6.0% (95% CI -10.5 to -1.6).
MASAI: AI screening meets its interval-cancer safety testJanuary 31, 2026. MASAI's interval-cancer endpoint met noninferiority, with specificity reported at 98.5% in both groups.

Read the dated findings and sources

Status: Measured screening safety; wider harms incomplete. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

January 31, 2026: MASAI's interval-cancer endpoint met noninferiority, with specificity reported at 98.5% in both groups. MASAI randomized screening trial follow-up.

Interpretation and tracking value

Noninferiority does not mean zero harm. This result does not quantify all false-positive procedures, overdiagnosis, or inappropriate treatments.

Next comparable measurement

Track recalls, benign biopsies, missed urgent cases, overdiagnosis estimates and treatment harms per eligible screened participant at fixed follow-up. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Patient benefit

15%estimated progress

What we track

Mortality, morbidity, functional outcomes, or quality of life at a specified follow-up, depending on the clinical problem.

Why this estimate?

Intermediate clinical findings suggest possible benefit, but mortality, quality-of-life and functional improvements are not established by the cited counts.

Advancements

COMPOSER sepsis model linked to 17% relative drop in sepsis deathsJanuary 23, 2024. Before-and-after study of 6,217 septic patients at two UC San Diego emergency departments: deployment of the COMPOSER deep learning model was associated with a 1.9% absolute (17% relative) reduction in in-hospital sepsis mortality (95% CI 0.3%-3.5%) and a 5.0% absolute rise in bundle compliance.
Randomized trial: AI-ECG alerts cut 90-day mortality to 3.6% from 4.3%April 29, 2024. Pragmatic randomized trial of 15,965 patients with 39 physicians in Taiwan: an AI-ECG alert identifying high-risk patients reduced 90-day all-cause mortality to 3.6% versus 4.3% in the control arm (HR 0.83, 95% CI 0.70-0.99).
CHARTwatch: non-palliative ward deaths 1.6% vs 2.1% after deploymentSeptember 15, 2024. Nonrandomized controlled study of 13,649 general internal medicine admissions at St. Michael's Hospital, Toronto: non-palliative deaths were 1.6% with the CHARTwatch early warning system versus 2.1% before (adjusted RR 0.74, 95% CI 0.55-1.00); the difference-in-differences estimate was not significant.
CONCERN trial: nursing-data AI warning lowers in-hospital death riskApril 2, 2025. Cluster-randomized trial of 74 hospital units and 60,893 encounters in two health systems: the CONCERN early warning system was associated with a 35.6% lower instantaneous risk of in-hospital death (adjusted HR 0.64, 95% CI 0.53-0.78) and 11.2% shorter length of stay; unanticipated ICU transfers rose 24.9%.
MASAI follow-up: fewer interval cancers with AI screeningJanuary 31, 2026. MASAI follow-up reported fewer invasive interval cancers descriptively: 75 versus 89.

Read the dated findings and sources

Status: Intermediate clinical evidence; survival/QoL effect unknown. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

January 31, 2026: MASAI follow-up reported fewer invasive interval cancers descriptively: 75 versus 89. MASAI randomized screening trial follow-up.

Interpretation and tracking value

Counts alone do not establish mortality, functional or quality-of-life benefit, and cannot be interpreted without group denominators and follow-up.

Next comparable measurement

Extend randomized follow-up to stage-specific morbidity, treatment burden, quality of life and mortality; distinguish exploratory from powered endpoints. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note