We conducted a quasi-experimental before-and-after study using historical controls to evaluate the impact of the AI-enabled emergency care system. The study was conducted across three counties (Hua, Neihuang, and Tangyin) in the Anyang region, Henan Province, China, representing a mixed urban-rural population of approximately 2.1 million residents. The pre-implementation period (January–December 2022) captured baseline performance with standard emergency care protocols. Following a phased implementation (January–June 2023), the post-implementation period (July–December 2023) evaluated system performance after full deployment.

The study population included all consecutive patients accessed by EMS for suspected SCD, ST-elevation myocardial infarction (STEMI), or other critical conditions during the study periods. Cases were identified from EMS dispatch logs and hospital emergency department records using ICD-10 codes. Inclusion criteria: (1) age ≥ 18 years; (2) transported by EMS to a participating hospital; (3) complete records of key variables. Exclusion criteria: (1) traumatic cardiac arrest; (2) declared dead before hospital arrival; (3) missing > 20% of core data elements. This study adheres to the STROBE guidelines for observational research (see STROBE checklist in Supplementary Table S3).

The Anyang Model system architecture rests on four integrated pillars (Fig. 1):

We developed and validated machine learning models following the guidelines by Luo et al. (PMID 27986644)⁷ using retrospective data from 5,247 emergency cases treated at the participating hospitals between 2019 and 2021.

Data source and cohort: Inclusion criteria: (1) age ≥ 18 years; (2) transported by EMS to the emergency department; (3) complete records of key variables (vital signs, ECG, outcomes). Exclusion criteria: (1) declared dead before hospital arrival; (2) traumatic cardiac arrest; (3) > 20% missing data.

The system was implemented in three phases:

Data were extracted from three sources: (1) EMS dispatch and run logs, (2) hospital electronic health records, and (3) system operation databases capturing algorithm outputs and platform usage metrics. Three primary outcome measures were pre-specified:

Secondary outcomes included AI alert accuracy during operation and protocol adherence rates.

Significant improvements were observed across all three primary outcome measures following system implementation (Table 1; Fig. 3).

IQR: interquartile range; STEMI: ST-elevation myocardial infarction; ROSC: return of spontaneous circulation; OHCA: out-of-hospital cardiac arrest.

The reduction in response time was observed consistently across all three counties and was more pronounced in rural areas (median reduction 4.2 min vs. 2.1 min in urban areas, p for interaction = 0.03; Fig. 4). The improvement in STEMI identification was partially attributable to the AI system: of the 32 additional correct identifications in the post-period, 12 (37.5%) were cases where EMS personnel initially did not suspect STEMI but the AI alert led to a secondary review and confirmation. The positive predictive value of the AI STEMI alert in operation was 78% (Supplementary Table S2). For OHCA, the absolute increase in ROSC from 18% to 31% represents a 72% relative improvement. In exploratory analyses, the benefit appeared concentrated among cases with witnessed arrest (ROSC 24% pre vs. 41% post, p = 0.02) and those with initial shockable rhythms (ROSC 31% pre vs. 52% post, p = 0.01).

The ITS analysis confirmed a significant level change in median response time immediately following implementation (β = −2.8 min, 95% CI: −3.7 to −1.9, p < 0.001), with no significant pre-existing trend (β = −0.05 min/month, p = 0.32). For ROSC, the ITS showed a level change of + 12% points (95% CI: +5 to + 19, p = 0.01). For STEMI identification, the ITS indicated a level change of + 22% points (95% CI: +13 to + 31, p < 0.01). Durbin-Watson tests showed no significant residual autocorrelation (response time: DW = 1.94, p = 0.58; ROSC: DW = 2.01, p = 0.82; STEMI ID: DW = 1.87, p = 0.41). Seasonal indicator variables were jointly non-significant (Wald test p > 0.20 for all outcomes). Adjusting for monthly average temperature and EMS staffing levels did not materially change these estimates. All three sensitivity ITS specifications yielded results consistent with the primary model (see Supplementary Tables S4–S5). These findings support that the observed improvements are temporally associated with the intervention rather than underlying secular trends.

During the 6-month post-implementation period, the AI system processed data from 621 emergency calls in real-time. The deterioration warning algorithm generated alerts for 84 cases (13.5% of all calls), with a positive predictive value of 41% for confirmed deterioration events. Of these 84 alerts, 50 were false positives, corresponding to a false positive rate of 8.1% (50/621 calls). The STEMI detection algorithm achieved 92% sensitivity and 96% specificity during operation, closely matching validation performance (Supplementary Table S2). Post-implementation surveys of EMS personnel (n = 112) indicated a median of 3 alerts per shift (IQR 1–5), and no respondent reported alert fatigue as a significant issue. Platform usage data indicated high adoption rates: 94% of EMS runs included at least one POCT device usage, and 89% of cases had complete data transmission from scene to receiving hospital before patient arrival.

All 47 ambulances across the three counties received identical hardware (IoT gateways, POCT devices, tablets). However, baseline connectivity differed: urban areas had 5G coverage during 98% of transmissions, while rural areas had 5G coverage during 82% (fallback to 4G, which did not affect data completeness). Uptake varied modestly by county: POCT usage was 96% in Hua County (urban core), 92% in Neihuang County (mixed), and 89% in Tangyin County (rural). AI alert acknowledgment rates were 91%, 87%, and 84%, respectively. Notably, the reduction in response time was largest in Tangyin County (4.2 min), suggesting that the system benefited rural areas most despite slightly lower fidelity. Detailed county-level data are provided in Supplementary Table S7.

The observed improvements reflect the synergistic effects of multiple system components. The 32% reduction in response time was multifactorial: (a) a dynamic vehicle assignment algorithm using real-time GPS and traffic data reduced dispatch-to-ambulance notification delays by an average of 45 s; (b) real-time traffic routing shortened en-route time by 68 s in urban areas; (c) parallel pre-arrival alerts to ambulances (on-board tablet) and hospitals eliminated sequential communication delays, saving an additional 30–40 s per call. The improvement in STEMI identification was driven by both the AI algorithm’s sensitivity (identifying 12 cases missed by EMS personnel) and systematic ECG transmission to receiving hospitals. The ROSC improvement — the most clinically significant finding — was associated with a cascade of process improvements: faster response enabling earlier CPR and defibrillation, better pre-hospital identification facilitating early activation of catheterization labs, and POCT enabling rapid treatment of reversible causes.

Notably, the benefits were more pronounced in rural areas (4.2 vs. 2.1 min reduction) and for witnessed arrests with shockable rhythms — precisely the subgroups where timely intervention has the greatest potential to alter outcomes. This suggests that integrated digital health systems may be particularly valuable in reducing urban-rural disparities in emergency care access.

Our findings align with and extend prior research on digital health interventions in emergency care. Studies of mobile health technologies have demonstrated improvements in response times and diagnostic accuracy⁶, and wearable devices have shown promise for cardiac monitoring⁸. However, most prior work has examined single technologies in controlled settings. The Anyang Model contributes to the literature by demonstrating that systematic integration of multiple technologies across the entire care continuum can achieve improvements that exceed the sum of individual parts. Our results support the World Heart Federation‘s call for health system-level digital transformation⁵.

Several mechanisms likely underlie the observed associations. First, the unified information platform breaks down information silos, ensuring that all stakeholders share a common operating picture. Second, real-time data transmission eliminates delays inherent in verbal handoffs and manual documentation. Third, AI decision support reduces cognitive load on EMS personnel, helping maintain high performance under stressful conditions. Fourth, the process re-engineering component ensures that technological capabilities are translated into consistent clinical actions through standardized workflows.

This study has several important limitations. First, the before-after design, despite ITS analysis, cannot prove causality. Unmeasured confounders (e.g., Hawthorne effect, concurrent training initiatives, seasonal variation in cardiac events, or simultaneous public CPR campaigns) may have contributed. We therefore frame our findings as demonstrating feasibility and strong temporal association, not system-level causal effectiveness. Second, the 6-month post-implementation follow-up is short. We acknowledge this as a preliminary report. The rationale for submission at this stage is to provide timely evidence for a rapidly evolving field and establish a baseline for future longitudinal follow-up (ongoing, with 24-month data to be reported separately). Third, the study was conducted in a single region of China, limiting generalizability to other settings with different healthcare systems and infrastructure. However, the three counties vary in urban-rural composition, suggesting relevance to other mixed urban-rural regions. Fourth, our primary outcome for SCD (ROSC) is an intermediate endpoint. Long-term follow-up data on survival to discharge and neurological status are being collected. Fifth, we did not conduct a formal cost-effectiveness analysis. Sixth, the AI algorithms were developed using local data; external validation in other regions is underway. Seventh, regarding generalizability to low- and middle-income countries (LMICs): while the full system requires broadband and devices, the modular architecture allows stepwise adoption. For example, the unified dispatch platform can be deployed first, followed by POCT, then AI. A less resource-intensive version using SMS-based alerts is under development. Eighth, the quasi-experimental design cannot exclude the possibility that unmeasured concurrent initiatives (e.g., community CPR training campaigns) contributed to the observed improvements, although ITS analysis reduces this concern.

Despite these limitations, our findings have several implications. First, digital health investments should prioritize system integration over standalone technologies — the whole is greater than the sum of its parts. Second, meaningful improvement is achievable even in resource-varying regional settings, challenging the assumption that advanced technologies require sophisticated urban academic centers. Third, technological implementation must be accompanied by workflow redesign and personnel training. For LMICs and underserved regions, the Anyang Model offers a potential blueprint with modular, phased adaptation.

Future research should include longer-term follow-up with survival and neurological outcomes, external validation in different geographic contexts, formal cost-effectiveness analysis, extension to other time-sensitive conditions (stroke, trauma, sepsis), integration with community-based interventions (public access defibrillation, CPR training apps), and continuous learning through feedback loops.

The authors declare no competing interests.

This study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Review Board of Hua County People‘s Hospital (Approval No. HC2021-089). The study involved analysis of anonymized system performance data and aggregated operational records, with no direct patient intervention for research purposes. The IRB waived the requirement for informed consent due to the retrospective and anonymized nature of the data and because the intervention was implemented as part of a health system quality improvement initiative. All data were de-identified prior to analysis.

Not applicable. This manuscript contains no individual person’s data.

The lead author affirms that this manuscript is an honest, accurate, and transparent account of the study; that no important aspects have been omitted; and that any discrepancies from the study as planned have been explained.

Data availabilityThe data that support the findings of this study are derived from a major science and technology project funded by the Anyang City Major Science and Technology Special Project (Grant No. 2025A02SF008). As the current work represents interim results of this ongoing project, the full dataset is subject to confidentiality requirements stipulated by the funding agency prior to the official project completion and public release. Therefore, the data are not publicly available at this time. Anonymized summary data and the analytical methods necessary to interpret, verify, and extend the findings are fully presented in the manuscript and supplementary materials. Requests for access to specific data may be considered on a case-by-case basis after the project’s official conclusion, and should be directed to the corresponding author.