Health AI research capacity and equity-adjusted patient value in 27 European Union countries: an ecological panel study, 2011–2024

Abstract

Background

Investments in health artificial intelligence (AI) are accelerating across European Union member states, yet evidence linking national AI research capacity to population-level health outcomes remains scarce. Most available evaluations focus on the performance of individual-level algorithms rather than system-wide effects. This study examines whether countries with greater health AI research activity achieve better population-level health value, and under what systemic conditions this relationship emerges.

Methods

We conducted an ecological panel study across all 27 EU member states for the period 2011–2024 (maximum 378 country-year observations). Two novel composite indices were constructed: the Equity-Adjusted Patient Value Index (EAPVI), aggregating self-perceived health, unmet medical needs, income-related equity gaps, and treatable mortality; and the Health AI Capacity Index (HAICI), capturing health AI publication volume, intensity, and international collaboration. Two-way fixed effects panel regressions with Arellano cluster-robust standard errors were estimated. Additional specifications included a HAICI×OOP (out-of-pocket expenditure) interaction, a COVID-19-era structural break model (HAICI×POST2020), Hansen threshold regression, and an event study design to validate parallel pre-trends. Fifteen robustness checks were conducted, including leave-one-country-out analysis, placebo tests, winsorization, and a spatial Durbin model.

Results

The Equity-Adjusted Patient Value Index (EAPVI) and the Health AI Capacity Index (HAICI) were not uniformly linked across the 27 European Union member states. Greater national health AI research capacity did not automatically translate into higher equity-adjusted patient value. The relationship between the two indices varied across financial-protection contexts. Marginal-effects analyses showed that higher HAICI was not consistently associated with more favourable EAPVI outcomes: the HAICI–EAPVI association was negative at lower OOP levels and became statistically indistinguishable from zero at higher OOP levels. This pattern is consistent with the multidimensional structure of EAPVI, in which component-level signals may move in opposite directions.

For treatable mortality, a marked shift emerged after the onset of the COVID-19 pandemic (P < .001). Countries with greater health AI research capacity experienced substantially fewer treatable deaths after 2020 — approximately 0.925 fewer per 100,000 population per index unit — with effects twice as large in Eastern compared to Western European countries. A Hansen threshold specification identified a HAICI threshold of 37.9, above which the HAICI–treatable mortality association became negative, although this result should be interpreted descriptively rather than as evidence of a causal threshold. The direction of these associations was supported by most robustness checks; however, a first-difference specification that accounts for the non-stationarity of HAICI produced a coefficient of opposite sign, indicating that the reported magnitudes are sensitive to the stationarity assumption.

Conclusions

Health AI research capacity does not automatically translate into improved population-level patient value. The observed associations were concentrated in country-years with lower out-of-pocket expenditure and were markedly stronger during the COVID-19 period than before it. This pattern is descriptive; the ecological design does not permit inferences about underlying causal mechanisms such as mobilization of research-side capacity during system stress. Because this is an ecological and correlational analysis, the findings do not identify specific policy levers. They indicate that in this panel the association between health AI research capacity and equity-adjusted population health outcomes was heterogeneous across financial-protection contexts and was most pronounced during the COVID-19 period, a pattern that merits further investigation with study designs capable of supporting causal inference.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12889-026-27676-8.

Background

The integration of artificial intelligence (AI) into health care has generated substantial expectations for improved diagnostics, treatment optimization, and system efficiency [1, 2]. In high-income countries, the expansion of research programs on health AI has been rapid, with the EU allocating considerable resources through the Horizon Europe program as well as national-level digital health strategies [3]. The evidence base for such investments is, however, limited, with most of the available information focusing on the validation of individual-level algorithms rather than system-level evaluations on whether countries that invest more in health AI also experience greater population-level outcomes [4]. Out-of-pocket (OOP) expenditure remains a significant barrier to equitable care access across EU member states [5], further complicating the translation of technological capacity into population-level gains.

Systematic reviews of AI in clinical settings document improvements in diagnostic accuracy across oncology, cardiology, ophthalmology [6–8], and infectious diseases, as well as gains in quality-adjusted life years and cost reductions attributable to fewer unnecessary procedures [9]. However, these evaluations are largely confined to single-institution or controlled trial settings, and their generalisability to population-level assessments remains uncertain. A separate and equally important distinction concerns the gap between research-side AI capacity and actual clinical implementation. A country that publishes extensively on health AI does not necessarily deploy the resulting algorithms, models, or decision support tools at scale in its hospitals and primary care system, since substantial translational, regulatory, infrastructural, and organisational barriers separate the publication frontier from the point of care [4]. Bibliometric indicators of health AI research therefore measure the upstream preconditions for clinical translation and the knowledge-producing capacity of a national system, not the extent to which AI is embedded in routine service delivery. Despite this research-to-deployment gap, there are both empirical and theoretical grounds for expecting national research-side AI capacity to leave a detectable footprint in population health indicators. At the individual level, controlled studies and systematic reviews have documented gains in diagnostic accuracy, earlier detection, and reductions in unnecessary procedures across multiple clinical domains [1, 2, 6–9], and these are precisely the mechanisms through which amenable and treatable mortality respond to improvements in clinical decision-making. At the system level, the diffusion of healthcare innovations is concentrated in settings with the scientific, professional, and infrastructural resources required to absorb and adapt them, so that countries with a richer health AI research base are typically also those best positioned to assimilate individual-level gains into routine care as the underlying technologies mature [4]. Population-level indicators such as treatable mortality, unmet medical needs, and income-related equity gaps therefore provide a plausible, albeit noisy, window through which the cumulative footprint of research-side capacity could in principle become visible, even though they remain far from a direct measurement of AI clinical deployment and although the magnitude of any such footprint must be expected to be modest in a fourteen-year panel. To our knowledge, cross-national comparative studies examining whether this research-side capacity is associated with equity-adjusted health system performance across a panel of EU member states remain limited, and the resulting evidence gap cannot be filled by algorithm-level evaluations alone.

The scale of the challenge facing EU health systems underscores the importance of this inquiry. More than 1.1 million deaths in the EU in 2022 were classified as avoidable, representing more than one in five deaths among EU citizens [10]. Stark East–West disparities persist: standardised treatable mortality rates in Eastern EU countries (120.5 per 100,000) were nearly three times higher than in Western Europe (44.7 per 100,000), with a more than sixfold difference between Latvia (316.8 per 100,000) and France (46.4 per 100,000) [10]. Unmet medical care needs display a similar gradient: while the EU average stood at 2.4% in 2023, rates reached 21.9% in Greece and in all member states the burden fell disproportionately on lower-income groups [11]. These inequalities suggest that the systemic conditions under which AI research capacity might translate into health gains are far from uniform across the EU.

This gap is consequential for health policy. Porter’s value-based health care framework defines value as health outcomes achieved per unit of cost, emphasizing that system performance should be judged by what patients experience [12]. If national AI research capacity translates into clinical and system-level improvements, we should observe its footprint in population health metrics, such as reduced amenable mortality, fewer unmet needs, and narrower equity gaps. Conversely, if AI research accumulates at the publication frontier without translating into clinical deployment at scale, no such association would emerge. An empirical test based on the association between research-side AI capacity and patient-value metrics is therefore a joint test of whether such capacity exists at a meaningful scale in each country and whether it reaches clinical practice with sufficient breadth to affect population health outcomes; the absence of an association cannot by itself distinguish between these two channels.

Measuring these relationships requires purpose-built indices. Existing health system performance measures typically focus on single outcomes or input-side indicators, making them ill-suited to capture the multidimensional construct of patient value or the research-intensive nature of AI capacity. To address this, we developed two novel composite indices: the Equity-Adjusted Patient Value Index (EAPVI), which integrates self-perceived health, unmet needs, equity, and treatable mortality into a single patient-centred metric; and the Health AI Capacity Index (HAICI), which captures the volume, intensity, and international integration of national health AI research programs. HAICI is therefore an indicator of research-side capacity rather than of clinical deployment or adoption, since it is constructed entirely from bibliometric data on peer-reviewed health AI publications and from the international collaboration network underlying those publications, and not from any direct measure of AI tool use in hospitals or clinics. Throughout the manuscript, HAICI should accordingly be read as a measure of a country’s capacity to generate and circulate health AI knowledge, and any association between HAICI and population-level outcomes reflects the cross-country correspondence between this research-side capacity and health system performance rather than a direct measurement of clinical AI uptake.

This inquiry sits at the intersection of three literatures that have until now engaged only loosely with one another. Health economics frames patient value and access through the lens of outcomes achieved per unit of cost and emphasises the central role of financial protection in translating system capacity into population-level welfare [5, 12]. Public health systems research locates the same question in terms of amenable and treatable mortality, unmet medical needs, and income-related equity gaps as markers of how effectively a system delivers care to the populations that need it [10, 11]. Innovation policy offers a third lens by distinguishing between the generation of new scientific knowledge and its absorption, adaptation, and diffusion within operational systems, and by noting that the gap between a national research frontier and routine clinical practice is conditional on organisational, regulatory, and financial features of the receiving system [4]. Our empirical design integrates these three perspectives in a single framework: HAICI operationalises the innovation-policy construct of research-side capacity, EAPVI operationalises patient value and equity in the public health systems tradition, and out-of-pocket expenditure enters as the health economics moderator that shapes whether and to what extent research-side capacity reaches the patient-experienced outcomes that the other two literatures emphasise.

We apply these indices to a panel of 27 EU member states over 2011–2024, using two-way fixed effects estimation to control for time-invariant country characteristics and common temporal shocks. Our analysis proceeds in two stages. First, we examine the unconditional relationship between HAICI and composite patient value. Second, because the COVID-19 period coincided with accelerated digital health adoption [13], we test whether the statistical association between HAICI and treatable mortality changed around 2020, without specifying or inferring a causal mechanism.

We emphasise at the outset that this is an ecological, correlational analysis. We present these results as descriptive evidence intended to inform the design of future studies and health AI investment strategies. Specifically, this study aims to: (1) examine whether national health AI research capacity, operationalised as the research-side capacity captured by HAICI, is associated with equity-adjusted patient value across EU-27 member states over 2011–2024; (2) test whether the statistical relationship between this research-side capacity and treatable mortality changed following the COVID-19 pandemic; and (3) assess whether out-of-pocket expenditure, as a demand-side financial-protection condition, moderates the association between research-side AI capacity and population health outcomes.

Methods

Data sources and sample

We assembled a country-year panel spanning all 27 EU member states for the period 2011–2024, yielding a maximum of 378 observations (27 countries × 14 years).

The 2011 lower bound was chosen because it is the earliest year in which the main source databases (Eurostat health statistics, Eurostat EU-SILC population survey, the World Bank World Development Indicators, and the OpenAlex bibliometric index) provide harmonised data for all 27 EU member states, allowing construction of a balanced-origin country-year panel without silent extrapolation over missing earlier observations. Earlier Eurostat EU-SILC rounds did not provide consistent coverage of income-stratified unmet medical needs across the full EU-27, and adopting an earlier start would therefore have introduced systematic gaps at the lower income quintile and in the equity-gap component of EAPVI. The 2024 upper bound reflects the most recent year for which OpenAlex bibliometric data and Eurostat demographic indicators were available at the time of analysis. Within this window, Eurostat treatable mortality and World Bank expenditure and workforce indicators are available through 2022 for most countries, which, combined with the one-year lagging of independent variables, produces the effective estimation samples of 311 observations for EAPVI and 312 for treatable mortality reported in the Methods section.

Data were drawn from three sources, and detailed variable definitions, database codes, and year coverage are reported in Table S11. Health outcome and system indicators; self-perceived health status, unmet medical needs due to cost, distance, or waiting, income-related inequalities in unmet needs (quintile 1 minus quintile 5 gap), and treatable mortality per 100,000 population under age 75; originated from Eurostat [14]. Health expenditure per capita, OOP spending as a share of current health expenditure, and physician density (physicians per 1000 population, based on the World Bank’s “practicing physicians” definition, SH.MED.PHYS.ZS) were obtained from the World Bank World Development Indicators [15]. Bibliometric data on health AI publications were extracted from OpenAlex [16], an open scholarly database, using Medical Subject Heading terms related to artificial intelligence, machine learning, deep learning, and natural language processing.

The OpenAlex search strategy combined two criteria. First, publications were required to be tagged with at least one of the following MeSH descriptors or OpenAlex topic equivalents: ‘Artificial Intelligence’, ‘Machine Learning’, ‘Deep Learning’, ‘Neural Networks, Computer’, ‘Supervised Machine Learning’, ‘Unsupervised Machine Learning’, ‘Natural Language Processing’, and ‘Computer Vision’. Second, publications were required to belong to the biomedical and health domain, operationalised as OpenAlex concept scores above 0.30 on at least one of the concepts ‘Medicine’, ‘Public Health’, ‘Health Care’, ‘Clinical Medicine’, ‘Epidemiology’, ‘Biomedical Engineering’, or ‘Health Informatics’. Inclusion was restricted to peer-reviewed journal articles and indexed conference proceedings published between 2011 and 2024 with at least one author affiliated with an institution located in an EU-27 member state, identified through the Research Organization Registry (ROR) records linked by OpenAlex. Editorials, commentaries, errata, retracted records, and pre-prints without peer review were excluded. Multi-country publications were counted once per country of affiliation for the international collaboration indicator and once for the publishing country’s total in the per-capita and share indicators. The query was executed on February 15–20, 2026 and validated by manual review of a random sample of 100 retrieved records per five-year bin, with a false-positive rate below 4% in all bins.

Because not all variables are available for every country-year (Eurostat treatable mortality data end in 2022 for some countries; World Bank indicators have sporadic gaps), the effective estimation samples are 311 observations for EAPVI models and 312 for treatable mortality models. This attrition from the theoretical maximum of 378 arises from three sources: missing values in health system covariates, one-year lagging of independent variables (which drops 2011 observations), and incomplete reporting by some smaller member states. All independent variables were lagged by one year to mitigate simultaneity. This study uses aggregate, publicly available data and does not involve patient-level research; accordingly, no patient engagement activities were undertaken.

Index construction

Equity-Adjusted Patient Value Index (EAPVI)

This index synthesizes four min–max normalized (0–100) components drawn from Eurostat. Self-perceived health is defined as the percentage of the population aged 16 years and over reporting their general health as ‘good’ or ‘very good’ on the five-point EU-SILC scale (very good, good, fair, bad, very bad), and enters as a positive indicator. Unmet medical needs is defined as the percentage of the population aged 16 years and over reporting at least one unmet need for medical examination or treatment during the previous 12 months due to cost, distance, or waiting time, and enters as a reversed indicator. The income-related equity gap is the percentage-point difference in reported unmet medical needs between the lowest (Q1) and highest (Q5) income quintiles and enters as a reversed indicator. Treatable mortality is defined as deaths before age 75 from causes that can mainly be avoided through effective and timely healthcare interventions, expressed per 100,000 population using the joint Eurostat–OECD treatable mortality list, and enters as a reversed indicator.

The formulation draws on Porter’s value-based health care framework [12] and the Dahlgren–Whitehead social determinants model [17]. Internal consistency assessed via Cronbach’s α [18] yielded α = 0.643 (standardized α = 0.662). This value falls below the conventional 0.70 threshold, which reflects the deliberate heterogeneity of the index: EAPVI is designed to capture distinct dimensions of patient value (subjective health, access, equity, clinical outcomes) rather than a single latent construct. Notably, the near-zero correlation between the equity gap and self-perceived health (r = − .012) indicates that these components provide complementary, non-redundant information. All four item-total correlations exceed 0.30 (range: 0.34–0.56), and leave-one-out analyses show correlations of 0.92–0.98 with the full index, supporting the inclusion of each component (Table S1).

Health AI Capacity Index (HAICI)

This index aggregates three equally weighted, min–max normalized bibliometric indicators: health AI publications per million population, the AI share of total biomedical research output, and international collaboration intensity (proportion of AI papers with foreign co-authors). Cronbach’s α was 0.831, indicating good internal consistency [18]. Convergent validity was confirmed through expected positive correlations with health expenditure per capita (r = .37) and negative correlations with treatable mortality (r = − .25) (Table S1).

The equal-weight, min–max formulation used for both indices was chosen for interpretability and for consistency with established composite indicator guidelines. To assess the sensitivity of our results to this weighting choice, we re-estimated HAICI using principal component analysis on its three underlying bibliometric indicators; the first principal component accounts for 75% of the joint variance and the resulting PCA-weighted index correlates with the original HAICI at r = .998 (Table S12), supporting the equal-weight formulation. For EAPVI, a single-factor solution is not the construct being measured: the index is designed to aggregate distinct dimensions of patient value, and its sub-threshold Cronbach α of 0.643 directly reflects this deliberate multidimensionality. Rather than impose a latent-factor structure on heterogeneous components, we report leave-one-component-out robustness (including an EAPVI-without-SPH specification in Table S4) and component-level regressions (Table S3), which allow the reader to see how each EAPVI component individually responds to HAICI. We regard this transparent component-level reporting as more informative than an EFA/CFA procedure, which would impose a homogeneity assumption that we explicitly reject in the construction of EAPVI.

Econometric strategy

Our identification strategy employs two-way fixed effects (TWFE) panel regression with country and year fixed effects, controlling for time-invariant unobserved heterogeneity and common temporal shocks [19, 20]. All standard errors are clustered at the country level using the Arellano estimator to address within-country serial correlation and heteroskedasticity [21]. We acknowledge the recent econometrics literature highlighting potential biases in TWFE estimators under heterogeneous treatment effects across cohorts and time [22, 23], and present complementary specifications (event study, threshold regression, first differences) to assess sensitivity to these concerns.

For EAPVI, we estimate three specifications. Model 1 includes only control variables: log health expenditure per capita (capturing overall resource availability), OOP share (financial protection), physician density (supply-side capacity), diabetes mortality rate (included as a confounder that proxies the underlying cardiometabolic disease burden of the population: diabetes is a major upstream risk factor for cardiovascular, cerebrovascular, and renal mortality, which are large components of treatable mortality, and variation in diabetes mortality across member states is substantial and largely independent of health AI research activity, so its inclusion reduces confounding from differential disease burden between high-HAICI and low-HAICI countries), and population aging (percentage aged 65 and above, capturing demographic pressure on health systems). Model 2 adds lagged HAICI. Model 3 introduces a HAICI×OOP interaction to test whether the association between research-side AI capacity and EAPVI varies across financial-protection contexts.

For treatable mortality, the baseline fixed-effects model (TM1) is augmented with a COVID-19 structural break specification that includes a HAICI×POST2020 interaction (TM-COVID). This specification tests whether the association between HAICI and treatable mortality differed during the pandemic period. The 2020–2022 period was characterised by a rapid and heterogeneous expansion of digital health across EU member states, documented in prior work [13]: telemedicine consultation volumes rose by more than an order of magnitude in several member states within weeks of national lockdowns; national and hospital-level deployments of AI-assisted triage, imaging diagnosis (particularly for chest radiography and chest CT), and outbreak-forecasting tools were accelerated under emergency procurement and regulatory fast-track schemes; and digital epidemiological infrastructure (case tracking, contact tracing applications, mortality surveillance dashboards) was expanded to scales that had not existed before the pandemic. These developments did not unfold uniformly across the EU-27, and their pace and depth plausibly correlated with the pre-existing research-side capacity captured by HAICI. The HAICI×POST2020 interaction is designed to detect whether the statistical association between research-side capacity and treatable mortality shifted in a manner consistent with this heterogeneous pandemic-era digital health expansion; the interaction term isolates a change in a statistical association rather than a causal activation of research-side capacity. Additionally, a Hansen threshold regression [24] endogenously identifies a critical HAICI value with bootstrapped confidence intervals.

An event study design decomposes the HAICI effect into year-specific interactions (reference: 2019) to validate parallel pre-trends. Robustness is assessed through leave-one-country-out analysis (LOCO), placebo tests with outcomes that should be unrelated to AI capacity (diabetes mortality, population aging), winsorized estimates (1st/99th percentiles), PCA-weighted index variants, country-specific linear trends, and a spatial Durbin model [25] with an inverse-distance geographic weight matrix. Prior to estimating the spatial Durbin model, we tested for spatial autocorrelation in the residuals of the baseline TWFE models using Moran’s I with an inverse-distance geographic weight matrix between country capitals, row-standardised. The test was computed separately for each cross-section of the balanced sub-panel (2014, 2016–2018) for both the EAPVI and the treatable mortality outcomes, with inference based on the analytical normal approximation and confirmed by 999 permutations. The spatial Durbin model was estimated on cross-sections for 2014 and 2016–2018; 2015 was excluded because physician density data were unavailable for three countries in that year, preventing construction of a balanced spatial weight matrix. Panel stationarity is evaluated via Im–Pesaran–Shin tests [26] and cross-sectional dependence via the Pesaran CD test [27].

All analyses were performed in R version 4.3.2 (R Core Team, 2023) using the plm package for panel estimation, the sandwich and lmtest packages for Arellano cluster-robust inference, the car package for variance inflation factor diagnostics, the psych package for internal-consistency statistics, and the spdep package for the spatial weight matrix, Moran’s I, and the spatial Durbin specification, with additional data manipulation in dplyr, tidyr, and readr and figures produced with ggplot2.

Results

Descriptive statistics

Table 1 presents summary statistics. Country-specific 2019 rankings for EAPVI, HAICI, treatable mortality, health expenditure, and OOP share are provided in Table S10. Mean EAPVI was 74.3 (SD = 13.8), ranging from 23.6 (Latvia, 2011) to 93.9 (Spain, 2019), reflecting substantial cross-country variation in patient value. Mean HAICI was 33.3 (SD = 17.0), with a minimum of 1.1 and maximum of 95.9, indicating wide dispersion in health AI research capacity across the EU-27. A 2019 country scatterplot of HAICI against EAPVI is shown in Figure S3. Mean treatable mortality was 194 per 100,000 under 75 (SD = 79.3). Average OOP expenditure constituted 21.0% of current health expenditure (SD = 9.4%). Health expenditure per capita averaged $3180 (SD = $2060), with a tenfold range from $402 to $8170 reflecting the economic heterogeneity between newer and older member states. Mean physician density was 2.08 per 1000 population (SD = 0.60), based on the World Bank’s practicing-physicians definition.

Table 1.

Descriptive statistics for key variables, EU-27, 2011–2024

Variable	N	Mean	SD	Min	P25	Median	P75	Max
EAPVI	350	74.3	13.8	23.6	67.5	76.2	86.2	93.9
HAICI	371	33.3	17.0	1.1	21.4	28.2	43.5	95.9
Treatable mortality	351	194	79.3	89.7	134	161	243	460
CHE per capita (USD)	366	3180	2060	402	1340	2550	5140	8170
OOP (% of CHE)	366	21.0	9.36	8.4	13.0	19.4	27.8	47.7
Physicians/1000	346	2.08	0.60	0.84	1.71	2.04	2.43	4.07
Diabetes mortality	351	26.6	16.0	8.9	16.7	22.0	30.9	115
Pop. 65+ (%)	378	18.9	2.70	11.6	17.6	19.2	20.6	24.6

EAPVI and HAICI on 0–100 scale. Treatable mortality = deaths per 100,000 under 75. Ns vary due to differential data availability

Abbreviations: CHE Current health expenditure, EAPVI Equity-Adjusted Patient Value Index, HAICI Health AI Capacity Index, OOP Out-of-pocket expenditure

Composite patient value

Table 2 reports the EAPVI regression results. In Model 1 (M1), the within R² of 0.081 indicates that the control variables explain a modest share of within-country EAPVI variation, reflecting the multidimensional nature of the index. Among controls, log health expenditure per capita showed a positive but nonsignificant association (5.735, P > .05), while OOP share was negative and nonsignificant (− 0.153, P > .05). Physician density was negatively associated with EAPVI (− 1.916), which may reflect reverse causation: countries with worse health outcomes may invest in expanding physician supply.

Table 2.

Two-way fixed effects panel regression: composite patient value (EAPVI)

	M1: Controls	M2: + HAICI	M3: + HAICI×OOP
HAICI (lag 1)	—	−0.082 (0.047)	−0.227** (0.072)
OOP (% CHE)	−0.153 (0.126)	−0.200 (0.108)	−0.397*** (0.114)
HAICI × OOP	—	—	0.007*** (0.002)
ln(CHE pc)	5.735 (4.194)	4.178 (4.694)	1.982 (4.673)
Physicians/1000	−1.916 (1.036)	−1.754 (0.928)	−2.012* (0.897)
Diabetes mort.	−0.000 (0.021)	−0.003 (0.023)	0.009 (0.024)
Pop. 65+ (%)	0.016 (0.853)	−0.064 (0.854)	0.091 (0.806)
Within R²	0.081	0.096	0.131
N	311	311	311

Country and year FE. Arellano cluster-robust SEs in parentheses. N = 311 reflects observations with complete data after one-year lagging

Abbreviations: CHE Current health expenditure, EAPVI Equity-Adjusted Patient Value Index, FE Fixed effects, HAICI Health AI Capacity Index, OOP Out-of-pocket expenditure

^*P < .05; ^**P < .01; ^***P < .001

In Model 2 (M2), the estimated association between lagged HAICI and EAPVI was negative but did not reach statistical significance at the conventional 5% threshold (β = −0.082, P = .086). Taken at face value, a one-unit increase in lagged HAICI corresponded to a 0.082-point reduction in EAPVI, but this point estimate is not statistically distinguishable from zero. After controlling for health expenditure, OOP, physician density, and demographic factors, higher AI research capacity was therefore not associated with a statistically significant improvement in composite patient value. This null-to-negative result is consistent with the interpretation that bibliometric AI capacity does not mechanically translate into patient-experienced value, at least not within the within-country variation captured by fixed effects.

The HAICI×OOP interaction in Model 3 (M3) was significant (β = 0.007, P < .001), revealing important heterogeneity. Marginal effects analysis (Table S7) showed that at the 10th percentile of OOP (11.3%), the marginal effect of HAICI on EAPVI was − 0.148 (P = .007), while at the 90th percentile (35.1%) it was effectively zero (0.020, P = .640). The negative sign at low OOP warrants careful interpretation: in countries with strong financial protection, increases in HAICI were associated with modest EAPVI reductions, possibly reflecting a composition effect whereby AI research expansion coincides with health system restructuring that temporarily disrupts value delivery. Alternatively, the composite nature of EAPVI may obscure opposing component-level effects: HAICI was significantly associated with increased unmet needs (β = 0.043, P = .043; Table S3), potentially reflecting that digitally advanced systems generate higher patient expectations, while showing a negative (though nonsignificant) association with treatable mortality.

Treatable mortality

Table 3 presents treatable mortality results. The baseline TWFE model (TM1) yielded a negative but nonsignificant HAICI coefficient (− 0.193, P = .261), consistent with no detectable unconditional association between AI capacity and treatable mortality.

Table 3.

Treatable mortality models: fixed effects, covid structural break, and threshold regression

	TM1: FE	TM-COVID	TM-Threshold
HAICI (lag 1)	−0.193 (0.172)	0.357** (0.135)	—
HAICI×POST2020	—	−1.282*** (0.232)	—
HAICI (below 37.9)	—	—	0.239 (0.199)
HAICI (above 37.9)	—	—	−0.273 (0.152)
OOP (% CHE)	−1.129 (0.704)	−1.831** (0.579)	−1.338* (0.544)
Physicians/1000	−5.680* (2.541)	−7.125** (2.466)	−6.978** (2.631)
Diabetes mort.	0.237* (0.111)	0.177 (0.108)	0.159 (0.114)
Pop. 65+ (%)	6.505* (2.541)	5.029* (2.194)	4.978* (2.477)
Within R²	0.060	0.163	0.091
N	312	312	312
Bootstrap P	—	—	< 0.001

Country and year FE. Arellano cluster-robust SEs. POST2020 = 1 for years ≥2020. Threshold: Hansen (1999) with 1000 bootstrap replications

Abbreviations: CHE Current health expenditure, FE Fixed effects, HAICI Health AI Capacity Index, OOP Out-of-pocket expenditure, TM Treatable mortality

^*P < .05; ^**P < .01; ^***P < .001

The COVID-era structural break specification (TM-COVID) revealed a more complex pattern. The HAICI level coefficient turned positive and significant (0.357, P = .009), while the HAICI×POST2020 interaction was strongly negative (− 1.282, P < .001). These coefficients must be interpreted jointly: the positive level effect indicates that, in the pre-2020 period, countries with higher HAICI had higher treatable mortality conditional on controls, likely reflecting the confounding role of health system modernization trajectories (countries actively building AI research infrastructure may simultaneously be addressing historically higher mortality burdens). The net effect of HAICI in the post-2020 period is the sum of these coefficients (0.357 − 1.282 = − 0.925), indicating that during the pandemic, a one-unit HAICI increase was associated with 0.925 fewer treatable deaths per 100,000. For a country at one SD above the mean HAICI (17 points), this translates to approximately 15.73 fewer treatable deaths per 100,000 after 2020.

The Hansen threshold regression identified an optimal HAICI threshold of 37.9 (bootstrap P < .001, F = 23.1). Below this threshold, the HAICI–mortality association was positive and nonsignificant (0.239, P = .229); above it, the association was negative and marginally significant (− 0.273, P = .074), suggesting that the HAICI–mortality association becomes negative only above this estimated threshold, although the above-threshold coefficient is only marginally significant. In the TM-COVID model, physician density showed the expected protective association (− 7.125, P < .01), while population aging was positively associated with treatable mortality (5.029, P < .05).

The event study confirmed parallel pre-trends (Fig. 1). Years 2012–2015 are pooled into a single pre-period indicator because one-year lagging drops 2011 observations and the remaining early-period years provide insufficient variation for individual-year estimation: pooling conserves degrees of freedom while still testing the pre-trend assumption. The joint Wald test that all pre-2020 HAICI×year interactions equal zero was not rejected (χ² = 3.47, df = 4, P = .483). Pre-2020 coefficients were small and statistically nonsignificant, whereas the 2020 interaction was − 1.21 (P = .026) and 2021 reached − 4.12 (P < .001). The effect attenuated after 2021 (2022: −0.24, P = .272; 2023+: 0.41, P = .247), suggesting a pandemic-period association pattern rather than a persistent post-pandemic association.

Fig. 1.

Event study: year-by-year HAICI×Period interaction on treatable mortality (Reference: 2019). 95% CIs from TWFE with HAICI×year interactions. Joint pre-trend test: χ² = 3.47, df = 4, P = .483. Reference period: 2019

Figure 2 displays marginal effects of HAICI on treatable mortality at different OOP levels. The protective effect was significant only at the 10th percentile of OOP (ME = − 0.44, P = .049) and declined monotonically, becoming positive (though nonsignificant) at the 90th percentile (ME = 0.18, P = .281). A similar gradient was observed for EAPVI (Fig. 2). These findings indicate that the negative association between HAICI and treatable mortality is concentrated in country-years with lower out-of-pocket expenditure; we do not interpret this as establishing a causal dependency on financial protection.

Fig. 2.

Marginal effect of HAICI on health outcomes across out-of-pocket expenditure levels (two panels). Delta-method SEs. Solid markers = significant (P < .05); hollow = not significant. Top panel: treatable mortality (TM-COVID model); bottom panel: EAPVI (Model 3)

Regional heterogeneity and confounding

The COVID-era HAICI effect on treatable mortality was approximately twice as large in Eastern EU (− 1.678, P = .004, n = 11 countries) as in Western EU (− 0.766, P = .002, n = 16) (Table S9 and Figure S2). The triple interaction HAICI×POST2020×EAST was − 1.522 (P < .001), confirming a statistically significant East–West differential (Table S9 and Figure S2).

This regional heterogeneity raises an important identification concern. Eastern EU countries have systematically higher OOP shares (mean 24.1% vs. 17.9% in the West) and lower baseline HAICI, creating potential confounding between the OOP interaction and the East–West structural divide. To disentangle these channels, we estimated a model stratified by initial HAICI level (median split at 2015 HAICI = 22.1) (Table S9). The triple interaction HAICI×POST2020×HIGH_HAICI was nonsignificant (0.645, P = .302), suggesting that the stronger Eastern EU effect reflects catch-up dynamics and higher marginal returns to AI capacity at lower baseline levels, rather than a simple artifact of OOP differences. However, we cannot fully separate these overlapping geographic and institutional patterns with 27 countries, and this remains a limitation.

Robustness and sensitivity

The direction of the COVID-era treatable mortality result was supported by 14 of the 15 robustness checks reported in Tables S5a and S12. The exception was the first-difference specification (0.481, P = .091), whose implications for the interpretation of the levels estimates are discussed as a key limitation below. Leave-one-country-out analysis: all 27 iterations yielded P < .001, with interaction coefficients ranging from − 1.35 to − 1.17 (Fig. 3). Placebo tests with diabetes mortality (P = .302) and population aging (P = .188) as dependent variables showed no significant HAICI×POST2020 effect, supporting specificity. Winsorized estimates (− 1.243, P < .001), country-specific linear trends (− 0.577, P = .012), and PCA-weighted HAICI (− 1.38, P < .001) all supported the main finding (Table S12). Moran’s I tests on the residuals of the baseline TWFE models (Table S13) indicated significant positive spatial autocorrelation in 2 of 8 cross-sections under analytical inference, concentrated in the treatable mortality outcome and in the earlier years of the balanced sub-panel (2014: I = 0.099, P = .005; 2016: I = 0.053, P = .040); this partial evidence of residual spatial structure motivated the spatial Durbin specification (direct effect: −0.170, P = .113; spillover: −0.990, P < .001; Table S6), which also supported the main finding. The nonsignificant direct effect but strongly significant spatial spillover in the Durbin model indicates that AI capacity in neighbouring countries is associated with improved domestic outcomes, consistent with cross-border knowledge diffusion. Maximum VIF was 1.83, indicating no multicollinearity concern.

Fig. 3.

Leave-one-country-out stability: HAICI×POST2020 on treatable mortality. All 27 estimates P < .001. Dashed line = full-sample estimate (− 1.282). Red markers = East EU; blue = West EU

In the HAICI component ablation (Table S2 and Figure S1), publications per million population () drove the effect (− 2.40, P < .001), while international collaboration alone was nonsignificant (− 0.37, P = .148). The wide 95% confidence interval for (− 3.39 to − 1.42) reflects the high cross-country variance in per-capita publication rates across the EU-27 (Figure S1), ranging from near-zero in smaller member states to over 90 in research-intensive countries; nevertheless, the interval excludes zero, confirming the robustness of this component’s contribution.

The standardized TM-COVID coefficients (Table S5b) show that the HAICI×POST2020 interaction (βstd = − 0.216) and OOP (βstd = − 0.221) are the two largest standardized effects among statistically significant predictors. Log health expenditure per capita has a larger absolute standardized coefficient (− 0.278) but does not reach significance (P = .065). The Pesaran CD test for cross-sectional dependence was nonsignificant for the EAPVI model (− 1.227, P = .220; Table S8), indicating that the Arellano cluster-robust standard errors provide adequate inference without requiring cross-sectional dependence corrections. Im–Pesaran–Shin tests confirmed that EAPVI and treatable mortality are stationary in levels (P = .021 and P < .001, respectively), while HAICI and health expenditure per capita are non-stationary, with first differences stationary for both (Table S8).

Discussion

To our knowledge, this is one of the first systematic, panel-econometric assessments of the relationship between national health AI research capacity and population-level patient value across the 27 EU member states. Using two novel composite indices designed to capture the multidimensional constructs of patient value and AI research capacity, we find results that are substantially more nuanced than a simple “AI improves health” narrative. We discuss both the supportive and cautionary findings, with particular attention to the methodological limitations that constrain causal interpretation.

Our findings align with, diverge from, and extend the existing literature in three distinct ways. They align with the body of clinical evidence documenting that individual AI-enabled tools can deliver measurable improvements in diagnostic accuracy, earlier detection, and reductions in unnecessary procedures [1, 2, 6–9], insofar as our pandemic-era treatable mortality finding is consistent with such gains becoming visible at the population level when deployment accelerates. They diverge from a simple reading of this literature that would predict an unconditional positive association between research-side AI capacity and patient value: our EAPVI models do not support such an unconditional association, and this divergence is consistent with the concerns raised by Kelly and colleagues about the gap between algorithm-level performance and clinical impact [4]. They extend the literature in two respects. First, they move the empirical base from single-institution or trial-level evaluations to a cross-national, panel-econometric framework, addressing the generalisability question that clinical reviews cannot answer by construction. Second, they show that any detectable footprint of research-side capacity on population outcomes is context-dependent, concentrated in settings with lower out-of-pocket expenditure and in the specific structural break of the COVID-19 period, rather than uniform across the EU-27. The telehealth transformation literature [13] is consistent with a pandemic-era change in the digital health environment within which research-side capacity could leave a detectable footprint, although, as discussed below, our design cannot establish the causal channel.

The composite patient value index showed no significant unconditional improvement with higher AI research capacity. This result is robust across multiple specifications (Table S4) and consistent with the view that bibliometric AI capacity captures research potential rather than clinical deployment.⁴ Countries may publish extensively on health AI without those innovations reaching clinical practice at scale. The component-level analysis (Table S3) further reveals that HAICI is associated with increased reported unmet medical needs (β = 0.043, P = .043), possibly because digitally advanced health systems generate higher patient expectations or because the measurement of unmet needs is more sensitive in systems with greater informational infrastructure. The composite EAPVI thus averages opposing signals across its components, yielding a null aggregate. This averaging effect is itself a likely consequence of the deliberate multidimensionality of EAPVI, reflected in its sub-threshold Cronbach α of 0.643: an index that intentionally captures distinct dimensions of patient value cannot be expected to aggregate into a single internally consistent latent factor, and the resulting averaging of heterogeneous component-level associations is a plausible direct source of the null aggregate findings in Models 2 and 3. Future research should examine these component-level dynamics, particularly the paradox of rising unmet needs in technologically advanced settings.

The HAICI×POST2020 interaction on treatable mortality is large, robust, and specific (passed both placebo tests). The event study pattern, with null effects pre-2020, a sharp shift in 2020–2021, and subsequent attenuation, is consistent with a pandemic-era change in the statistical association rather than a pre-existing trend. One descriptive interpretation consistent with this pattern is that countries with greater health AI research infrastructure also experienced wider or earlier deployment of telemedicine, AI-assisted triage, and data-driven public health responses during the acute pandemic period [13]; however, our data cannot distinguish this channel from alternative explanations, including correlated institutional reforms, differential pandemic exposure, and the reverse feedback from treatable mortality to subsequent HAICI documented in our Granger analysis (β = −0.063, P = .027). The considerations below further temper any causal interpretation. The considerations below further temper any causal interpretation.

First, the positive HAICI level coefficient in the TM-COVID model (0.357, P = .009) indicates that before the pandemic, higher HAICI was associated with higher treatable mortality after controlling for expenditure, physician supply, and demographics. This counterintuitive finding likely reflects selection: countries building AI capacity may have been those with the greatest room for improvement in treatable mortality, representing a form of regression-toward-the-mean pattern that the interaction specification must be interpreted against.

Second, panel unit root tests show that HAICI is non-stationary in levels (Im–Pesaran–Shin statistic = 9.64, P = 1.000), although first differences are stationary (P < .001) [26]. The treatable mortality dependent variable is I(0). This asymmetry raises the possibility of spurious regression in levels specifications. When we estimated a first-difference model, the HAICI coefficient reversed sign (0.481, P = .091), though the small and trending sample makes first-difference estimation imprecise with only 14 years of data. We retain the levels specification as our baseline because treatable mortality is stationary and the fixed-effects estimator is generally robust to I(1) regressors when the dependent variable is I(0) [20], but the non-stationarity of HAICI is a meaningful limitation that we flag explicitly in the Limitations section below.

Third, Granger causality tests revealed no significant forward effect of HAICI on treatable mortality (χ² = 0.017, P = .897), but significant reverse feedback from treatable mortality to HAICI (β = −0.063, P = .027). This suggests that countries experiencing declining treatable mortality subsequently invest more in health AI research, consistent with a resource-availability mechanism: health system improvements free up resources and institutional capacity for research investments. The causal arrow may thus run partly from health outcomes to AI capacity, not only the reverse. This reverse channel does not invalidate the interpretation of the COVID-era interaction as a change in statistical association, but it underscores that the unconditional levels relationship is endogenous.

The HAICI×OOP interaction is directionally consistent across both outcomes: the negative association of HAICI with treatable mortality and the less negative (or null) association of HAICI with EAPVI are concentrated in settings with low OOP expenditure. This pattern is descriptively aligned with the broader health economics literature on supply-side technological improvements and demand-side financial protection [5], but our ecological design cannot establish that financial protection is a causal precondition for AI-related health gains. However, the geographic concentration of low-OOP countries in Western and Northern Europe means that the OOP interaction partially captures East–West structural differences. Our stratification analysis suggests this is not a pure artifact, but the confounding cannot be fully resolved with 27 cross-sectional units.

Recent work on staggered difference-in-differences designs has shown that conventional TWFE estimators can produce biased estimates when treatment effects are heterogeneous across cohorts and time [22, 23]. In our setting, HAICI is a continuous treatment rather than a binary intervention, and countries “adopt” AI research capacity gradually rather than at discrete treatment times. The Goodman-Bacon decomposition is therefore not directly applicable, but the concern about heterogeneous effects remains relevant. Our event study, which estimates year-specific effects, partially addresses this by allowing the HAICI–mortality relationship to vary freely over time. The null pre-2020 effects and the sharp post-2020 break are reassuring, but we cannot rule out that the interaction captures differential COVID mortality shocks correlated with, but not caused by, AI capacity.

Limitations

First, EAPVI’s Cronbach α of 0.643 falls below the conventional 0.70 threshold. Although this reflects deliberate component heterogeneity, it means the index captures multidimensional patient value at the cost of internal consistency. The component-level analysis in Table S3 shows that the individual EAPVI components can respond in opposing directions to HAICI, with unmet medical needs increasing (β = 0.043, P = .043) while treatable mortality declines, so the composite index averages opposing signals. The sub-threshold alpha is therefore directly linked to, and may be partly responsible for, the null aggregate EAPVI findings reported in Models 2 and 3. This limitation is intrinsic to any value-based composite that privileges construct breadth over internal homogeneity. Second, HAICI measures research capacity through publications, not clinical AI deployment; countries may have high HAICI scores without widespread clinical AI use, and vice versa.

Third, the study is ecological in design, and the observed country-level associations between HAICI and treatable mortality cannot be translated into claims about the clinical deployment of specific AI tools in local health systems. High national publication output does not guarantee that the algorithms, models, or digital health solutions generated by that research reach the point of care within the same country. Accordingly, HAICI should be interpreted as an indicator of national research intensity rather than of local AI implementation, and any inference from these aggregate patterns to patient-level or clinic-level mechanisms would risk the ecological fallacy.

Fourth, HAICI is non-stationary in levels (IPS statistic = 9.64, P = 1.000) whereas treatable mortality is stationary (P < .001). A first-difference specification that accounts for this asymmetry produces a positive HAICI coefficient of 0.481 (P = .091), opposite in sign to the levels estimate. Although the short time dimension of the panel (14 years) limits the precision of first-difference estimation and the fixed-effects estimator is generally robust to I(1) regressors when the dependent variable is I(0) [20], this sign reversal indicates that the quantitative magnitudes reported for the COVID-era treatable mortality association are sensitive to the stationarity assumption, and the primary findings should be read as robust in direction but uncertain in magnitude.

Fifth, HAICI is constructed from bibliometric data extracted from OpenAlex, which introduces several sources of measurement error that cannot be fully eliminated. MeSH-based topic tagging and OpenAlex concept scores rely on automated classification of article metadata and abstracts, which can misclassify papers that use AI methods without foregrounding them in titles or keywords, or conversely over-include papers that mention AI in passing. Language coverage in OpenAlex is stronger for English than for other European languages, which may understate the output of member states with large domestic-language scholarly traditions. The international collaboration indicator depends on author-affiliation parsing through Research Organization Registry records, which are more complete for large universities than for hospital-based research groups. These measurement errors are unlikely to be systematically correlated with EAPVI or treatable mortality and are therefore more likely to attenuate than to invert the estimated associations, but they introduce an additional source of attenuation bias that is not reflected in the reported standard errors.

Sixth, the interval between the appearance of health AI research at the publication frontier and its eventual footprint on population-level outcomes is unlikely to be a single year. New clinical AI tools typically require several years for regulatory clearance, pilot deployment, integration into institutional workflows, clinician training, and scale-up before they can affect amenable or treatable mortality at the population level. Our one-year lag on independent variables is chosen primarily to mitigate simultaneity rather than to capture the full translation horizon, and a two-year lag sensitivity analysis (Table S4 and Table S5a) produces coefficients in the same direction but with attenuated magnitudes. A panel of only fourteen years does not permit reliable estimation of longer distributed-lag structures, and it is possible that the full impact of research-side capacity on population outcomes would become detectable only over a longer horizon than the one covered here. This is a fundamental constraint of any analysis that combines fast-moving bibliometric indicators with slow-moving health outcomes in a short time dimension.

Conclusions

The study shows that health AI research capacity was not systematically associated with stronger patient value at the population level. The central result is that the link between AI and health system performance is shaped by context: AI-related research strength should therefore be interpreted in relation to the health-system and financial-protection contexts in which its association with population-level outcomes is observed.

This broader conclusion is reflected in the relationship between the two indices developed in the study. HAICI captures national health AI research capacity, whereas EAPVI captures patient value through a multidimensional combination of outcomes, access, and equity. The fact that higher HAICI does not systematically correspond to higher EAPVI suggests that research intensity alone is not enough to generate broad patient-centered value across health systems.

At the same time, the association between HAICI and treatable mortality was markedly stronger during the COVID-19 period than before it. This statistical pattern is descriptive and does not by itself demonstrate that AI research capacity was causally mobilized during the pandemic. Alternative explanations that our design cannot rule out include correlated institutional reforms, heterogeneous pandemic exposure across member states, and the reverse feedback from declining treatable mortality to subsequent HAICI identified in our Granger analysis.

Overall, the association between HAICI and EAPVI in the EU-27 panel was heterogeneous rather than uniform. Within this sample, the negative associations between HAICI and treatable mortality were concentrated in country-years with lower out-of-pocket expenditure and in the period after 2020, while the unconditional association with composite patient value was not positive. Together with the reverse feedback documented in the Granger analysis, these findings indicate that the cross-country relationship between research-side AI capacity and population-level patient value is context-dependent and bidirectional. Interpretations of these patterns as direct policy prescriptions should await study designs with stronger identification than the ecological panel used here.

Abbreviations

EAPVI

Equity-Adjusted Patient Value Index

HAICI

Health AI Capacity Index

CHE

Current health expenditure

CD

Cross-sectional dependence

FE

Fixed effects

LOCO

Leave-one-country-out

ME

Marginal effect

OOP

Out-of-pocket expenditure

PCA

Principal component analysis

SD

Standard deviation

SE

Standard error

TM

Treatable mortality

TWFE

Two-way fixed effects

Authors’ contributions

HI: Conceptualization, Methodology, Formal Analysis, Writing – Original Draft; MAB: Methodology, Analysed the data, Writing – Original Draft; ICI: developed the composite indices, Methodology, Writing – Original Draft; OA: Methodology, Analysed the data, Writing – Original Draft. All the authors reviewed the final version of the manuscript before submission. The authors approved the final manuscript. All authors reviewed the manuscript.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Data availability

All data used in this study are publicly available. Health outcome and health system indicators were obtained from Eurostat (https://ec.europa.eu/eurostat), health expenditure and physician density data from the World Bank World Development Indicators (https://data.worldbank.org), and bibliometric data on health AI publications from OpenAlex (https://openalex.org). The datasets and analysis code supporting the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

This study is based exclusively on aggregate, publicly available data obtained from Eurostat, the World Bank World Development Indicators, and OpenAlex. No individual-level, patient-level, or identifiable data were used. In accordance with applicable national and institutional regulations, ethical approval and consent to participate were not required for this type of research.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.