AI-Assisted Village Doctors and Rural Primary Healthcare: A Randomized Controlled Trial in China

Village doctors in treatment villages receive standardized training on an AI health assistant app ("Ant AFu," developed by Ant Group) and supported access to the app for use in their daily clinical work. Training covers the app's core functions, guidance on when AI input is and is not appropriate, how to incorporate the tool into diagnosis, prescribing, chronic-disease follow-up, and home visits, and how to recognize high-risk cases requiring referral. Training is followed by hands-on practice with simulated and common real-world cases, and by several weeks of follow-up support (group chat/phone-based Q&A and periodic check-ins) after initial rollout, plus periodic refresher sessions and targeted outreach to low-usage doctors during the intervention period. There are two treatment arms: in one, only village doctors receive training and app access; in the other, villagers in the same villages are also encouraged and supported to use the app themselves. Villages in the two control arms continue current practice during the study period and are offered the same training after the endline survey (delayed treatment).

2026-08-12

2027-09-30

(1) Service coverage: monthly number of unique patients seen per doctor, total monthly patient volume, monthly chronic-disease follow-up visits completed. (2) Quality of diagnosis and prescribing: a standardized diagnostic/prescribing quality score derived from (a) audited case records for tracer conditions, (b) doctor performance on standardized clinical vignettes, and (c) doctor behavior recorded during unannounced standardized patient (SP) visits. (3) Patient recovery: symptom improvement, self-reported recovery, repeat visits/ hospitalization, and adherence to prescribed treatment, measured via phone follow-up at fixed intervals after the index visit (e.g., day 7/14 for acute conditions, day 30/90 for chronic conditions).

Service coverage outcomes are drawn directly from monthly case-record/ledger data collected from each village clinic; no construction is needed beyond aggregation to the doctor-month or village-month level. The diagnostic/prescribing quality index is constructed for a small set of pre-specified tracer conditions (to be finalized from baseline ledger review; likely candidates include acute respiratory infection/fever, gastrointestinal symptoms, hypertension follow-up, diabetes follow-up, minor skin/wound infections, and chronic-disease medication management for the elderly). For each sampled case (from record audits, vignettes, and SP encounters), we score: completeness of history-taking, whether danger signs were asked about, whether necessary physical exams were performed, appropriateness of the prescription, absence of inappropriate antibiotic use, absence of inappropriate steroid/IV use, appropriateness of referral decisions, and clarity of medication instructions. These binary/ordinal sub-items are standardized and averaged (or aggregated via a weighted index, method to be pre-specified in the pre-analysis plan) into a single quality index per case, then averaged to the doctor level. The patient recovery outcome is constructed from a structured phone follow-up instrument covering symptom resolution, self-rated health, recurrence/repeat visits, hospitalization/ER use, complications, medication adherence, and satisfaction; these items will be combined into a summary recovery index (standardized and averaged, per the pre-analysis plan), with key individual items also reported separately.

(1) Depth/quality of home visits and follow-up (a constructed home-visit-depth index). (2) Patient costs and time use: total visit cost, drug cost, referral/travel cost, time lost from work, and time from symptom onset to treatment. (3) Doctor knowledge, confidence, and behavior change: vignette test scores, case judgment accuracy, self-reported confidence, willingness to conduct follow-up visits or take on more complex cases. (4) Safety indicators: rate of clearly inappropriate prescriptions, antibiotic/steroid overuse rates, missed danger-sign rate, delayed-referral rate, and patient adverse events. (5) Mechanism/usage indicators: app activation, weekly/monthly active use, number of queries per session, functions used (symptom check, report interpretation, follow-up support, health education), and whether promotion to village doctors generates spillover use among villagers (in Treatment 2 villages).

The home-visit-depth index is constructed from a checklist administered to sampled patients/households: whether a home visit occurred, whether vital signs were measured, whether medication adherence was checked, whether prior history was reviewed, whether danger signs were explained, whether follow-up was scheduled, whether records were kept, whether caregivers were contacted, visit duration, and visit frequency; sub-items are standardized and averaged into a single index. Cost and time-use outcomes are collected directly from patient/household follow-up surveys and require no construction beyond standard aggregation (e.g., total cost = consultation + drug + travel cost). The safety index aggregates the inappropriate-prescribing indicators listed above (antibiotic overuse, steroid/IV overuse, missed danger signs, delayed referral, adverse events) into a single standardized safety score, with each component also reported individually given their policy relevance. Usage/mechanism variables are drawn from de-identified backend log data provided by Ant Group (login frequency, session counts, function-level usage, query-type tags) and are used primarily for mechanism analysis (e.g., dose-response/TOT analysis, exploring what predicts effective use) rather than as headline treatment-effect outcomes.

This is a village-level cluster-randomized controlled trial conducted in a mountainous county in Guizhou Province, China, covering all 140 administrative villages in the county. After a baseline survey of village doctors and clinics, villages are randomly assigned within townships (stratified randomization) to one of four arms of 35 villages each: (1) Treatment 1, in which village doctors receive standardized training and supported access to an AI health assistant app; (2) Treatment 2, which adds village-level promotion of the app to residents; and (3)-(4) two control arms that continue existing practice during the study period, with the same training offered after the study ends. Randomization is stratified within townships to balance village population, doctor gender, and doctor education. The intervention period lasts approximately 6 months, with a baseline survey beforehand and an endline survey afterward (total study duration approximately 12-13 months). Outcomes are measured through doctor surveys, monthly case-record collection, sampled patient phone follow-up, household surveys, and standardized clinical assessment methods, with analysis conducted at the village level using intention-to-treat comparisons and village-clustered standard errors.

Not available

Randomization was conducted in office using a computer-generated randomization procedure (stratified randomization within townships), implemented after completion of the baseline village/doctor survey. Balance across arms was verified using standardized difference tests on stratification and key baseline covariates.

Administrative village (cluster). All doctors and patients within a given village clinic receive the same treatment assignment as their village. There is a single level of randomization (village-level, stratified within townships); the post-intervention access-withdrawal sub-study introduces a second, later-stage individual (doctor-level) randomization within the pool of previously treated doctors.

Yes

140 villages

Approximately 210 village doctors (1-2 doctors per village, 1.5 average across 140 villages) and approximately 1,400 patients (10 average across 140 villages sampled per village for follow-up), plus village-level administrative records and de-identified app usage data covering all 140 villages.

35 villages - Control Arm 1 (existing practice maintained; delayed training after endline)
35 villages - Control Arm 2 (existing practice maintained; delayed training after endline)
35 villages - Treatment 1 (village doctor training + AI app access)
35 villages - Treatment 2 (village doctor training + AI app access + villager-level promotion of app use)

Minimum Detectable Effect Size for Main Outcomes