Collaborating on AI research helps medtech firms get AI devices past regulators and into use—small firms benefit most— and those that succeed earn a persistent labour-productivity premium though margin gains are limited and fragile.
AI Innovation and Firm Performance in the Medical Device Industry
arxiv
quasi_experimental
medium evidence
8/10 relevance
Full text
usable extracted full text
Source
PDF
Using a novel dataset linking FDA AI-device clearances, patents, publications, and firm accounts, the paper shows that external AI research collaboration raises the likelihood a medtech firm brings AI devices to market (especially for small firms), and that introducing AI devices is associated with higher labour productivity but weaker, less-robust effects on profit margins.
Citation observations
Cumulative provider counts captured on specific dates; providers are never combined.
No provider observation is available for this paper.
Missing data, not a zero citation count.
Whether artificial intelligence pays off for the firms that build it into their products is hard to establish, because AI innovation is itself hard to observe. The medical technology sector is a rare exception: an AI-enabled device must obtain clearance from a national health authority before it can reach a patient, leaving a dated, firm-attributable record of AI innovation output that can be observed directly rather than proxied. We exploit this setting with a three-stage recursive model estimated on a novel firm-level dataset linking FDA premarket clearances, USPTO patents, Scopus publications, and Orbis financials, tracing the full innovation chain from external collaboration through AI device introduction to firm performance. We find that external AI research collaboration is a robust driver of AI device introduction across firm sizes and estimators, with a larger effect for small firms, consistent with external knowledge ties substituting for limited internal R&D capacity. Decomposing by partner type, the effect is largest for industry and clinical collaborations and smallest for academic ties, consistent with the former being closer to the regulatory and commercialisation process. Firms that bring AI devices to market display higher labour productivity, an effect robust for small firms and the full sample that holds under both sequential and joint maximum-likelihood estimation and accumulates across successive device introductions. Effects on profit margins are present but weaker and do not survive all specifications, a pattern consistent with competitive entry eroding pricing power as AI devices diffuse through the sector.
Summary
Main Finding
External AI research collaboration materially increases the likelihood that a medtech firm brings an AI-enabled device to market, especially for small firms, and firms that introduce AI devices exhibit a robust labour-productivity premium (which accumulates with successive device introductions). Effects on profit margins are weaker and not robust across specifications, consistent with competitive diffusion eroding pricing power.
Key Points
- Novel direct-output measurement: the authors exploit FDA premarket clearances to observe AI innovation output (AI-enabled devices) rather than relying on proxies (jobs, resumes, generic patents).
- Three-stage recursive identification links inputs → outputs → performance:
- External AI collaboration (instrumented)
- AI device introduction (binary and cumulative counts, instrumented)
- Firm performance (labour productivity and profit margins)
- Instruments:
- For collaboration intensity: firm AI publication count and geographic distance to nearest AI research hub.
- For device introduction: AI-specific patent flow built from the AIPD.
- Collaboration raises the probability of introducing AI devices; the marginal effect is larger for small firms (consistent with substitution for limited internal R&D).
- Partner-type decomposition: industry and clinical collaborations have the largest impact on device introduction; academic ties have the smallest effect (suggesting proximity to regulatory/commercial channels matters).
- Performance effects:
- Labour productivity: positive, robust across estimators and both sequential and joint maximum-likelihood estimation; accumulates with multiple device introductions.
- Profit margins: positive but weaker and sensitive to specification — do not survive all robustness checks.
- Regulatory context: ~94% of FDA-listed AI devices entered via the 510(k) pathway; Radiology, Cardiovascular, and Neurology account for >90% of AI device approvals in the sample.
- Robustness: results hold across several estimators and specification choices; the authors explicitly link collaboration-driven innovation to outcomes to mitigate confounding by firm heterogeneity.
Data & Methods
- Data sources linked at firm level:
- FDA premarket clearance records for AI-enabled and all 510(k) devices (FDA list retrieved Dec 2025; 1,247 AI devices; 125,678 total 510(k) devices).
- USPTO patent data via PatentsView and the Artificial Intelligence Patent Dataset (AIPD).
- Scopus scientific publication records (AI-related publications).
- Orbis (Bureau van Dijk) financials for firm-level outcomes (matched to 5,652 firms; sample covers firms active in Radiology, Cardiovascular, Neurology).
- Sample focus: three specialties (Radiology, Cardiovascular pathologies, Neurology) that represent >90% of FDA-cleared AI devices.
- Empirical design:
- Three-stage recursive model (akin to a CDM framework) to sequentially link collaboration → device introduction → performance.
- First stage: external collaboration intensity instrumented by AI publications and distance to AI research hubs.
- Second stage: fitted collaboration predicts AI device introductions; device introduction instrumented by AI-specific patent flow.
- Third stage: predicted device introduction tied to labour productivity and profit margin outcomes.
- Estimation performed both sequentially and via joint maximum-likelihood; robustness checks include firm-size splits and partner-type decompositions.
- Key sample facts: ~30,886 devices (in the three specialties) filed by 7,290 applicants; final matched financial sample ~5,652 firms (78% match rate to Orbis).
Implications for AI Economics
- Measurement: Regulatory approvals are a valuable, underused direct measure of AI product innovation. Where available, device/approval registries let researchers observe realized AI-enabled products rather than inputs or intentions.
- Role of external knowledge and absorptive capacity: External research ties (especially clinical and industry partners) are central to translating AI research into regulated products. Small firms disproportionately benefit from such ties, highlighting the substitutability of external knowledge when internal R&D capacity is limited.
- Heterogeneous returns: AI-driven productivity gains can be distinct from price/profit gains. Productivity gains appear durable and cumulative, while margin effects are fragile and likely dissipate with diffusion and competitive entry. Models of AI adoption should therefore distinguish between efficiency improvements and appropriation of returns.
- Partner-type and geography matter: Collaboration with actors close to regulatory and commercial processes (industry, clinical) more effectively converts AI knowledge into marketable devices than academic-only ties. Geographic proximity to AI hubs facilitates collaboration—implying localized ecosystem effects.
- Policy and market design:
- Policies that lower frictions to clinical and industry collaboration (data access, safe-harbour frameworks for data sharing, funding for translational partnerships) can speed commercialization, especially for smaller firms.
- Regulatory pathways and timing affect innovation incentives: the predominance of 510(k) clearances suggests incremental pathways dominate current AI device commercialization; policy changes to pathways or exclusivity could reshape firm incentives.
- Given weaker margin effects, policymakers should monitor competition dynamics as AI devices diffuse; diffusion may increase social welfare via productivity but compress firm rents.
- Research directions:
- Extend the regulatory-approval approach to other regulated AI applications (e.g., in vitro diagnostics, digital therapeutics, or non-health regulated markets) to test generalizability.
- Explore mechanisms behind weaker margin effects: pricing competition, reimbursement pathways, or regulatory-driven standardization.
- Investigate long-run dynamics of data accumulation and incumbent advantage in AI-enabled healthcare products.
Limitations to note (explicit in the paper): focus on medtech specialties dominating FDA AI approvals may limit generalizability to other sectors; identification relies on instrument validity (distance to hubs, publications, AI patents) which plausibly reduce but do not wholly eliminate endogeneity concerns; Orbis match attrition and newer firms may be underrepresented.
Assessment
Paper Typequasi_experimental
Evidence Strengthmedium — The paper leverages a novel, direct outcome measure (FDA-cleared AI-enabled devices) and a coherent recursive IV structure that links inputs, outputs, and outcomes, which substantially improves over proxy-based studies. However, causal claims rest on instrument exclusion restrictions that are plausible but not ironclad (publication counts and distance may correlate with unobserved productivity determinants; patent flows may directly relate to performance), and the setting is sector- and regulation-specific, limiting external validity.
Methods Rigorhigh — The authors construct a rich, linked dataset (FDA approvals, USPTO patents, Scopus publications, Orbis financials), use a multi-stage recursive econometric framework appropriate for the innovation–performance chain, implement IVs with theoretically motivated instruments, test alternative estimators (sequential vs joint MLE), and decompose effects by partner type and firm size; remaining concerns center on instrument validity and remaining selection/omitted-variable threats.
SampleFirm-level panel linking: FDA AI-enabled device list (1,247 AI devices as of Dec 2025) and full 510(k) premarket approvals (125,678 devices total), focusing on three specialties (Radiology: 956 AI devices; Cardiovascular: 116; Neurology: 56) which together account for >90% of AI device approvals; universe of firms with at least one premarket-approved device in these specialties (30,886 devices from 7,290 applicants), matched to Orbis financials yielding 5,652 firms (78% match rate); USPTO patent data via PatentsView and AI-specific patent flows from the AIPD; Scopus publication counts and measures of external collaboration intensity and partner type; snapshot period up to December 2025.
Themesproductivity innovation human_ai_collab
IdentificationThree-stage recursive IV model: (1) instrument external AI collaboration intensity with firm AI publication counts and geographic distance to the nearest AI research hub; (2) instrument AI device introduction (binary and cumulative counts) with firm AI-specific patent flow from the AIPD; (3) link predicted AI device introduction to firm outcomes (labour productivity, profit margins); robustness checked via sequential and joint maximum-likelihood estimation and decomposition by partner type and firm size.
GeneralizabilityFocus on regulated medical devices; results may not apply to non-medical sectors, US FDA-based identification (and 510(k)-dominant pathway) biases sample toward lower-risk/cleared devices and US regulatory context, Analysis restricted to three specialties (Radiology, Cardiovascular, Neurology) that dominate AI-device approvals, Matched Orbis financials exclude some very small or very new firms, limiting inference for micro-startups, IV-based identification requires strong exclusion restrictions which may not hold in other contexts or geographies
Claims (11)
| Claim | Direction | Outcome | Confidence & Evidence | Details |
|---|---|---|---|---|
| External AI research collaboration significantly increases the probability that a medical technology firm introduces an AI-enabled medical device. Adoption Rate | positive | Probability of introducing at least one AI-enabled medical device |
Reading fidelity
high
Study strength
medium
|
not reported
|
| The positive effect of external AI collaboration on AI device introduction is larger for small firms. Adoption Rate | positive | AI-enabled device introduction, with heterogeneous effects by firm size |
Reading fidelity
high
Study strength
medium
|
not reported
|
| The effect of external collaboration on AI device introduction is largest for industry and clinical partners and smallest for academic partners. Adoption Rate | positive | AI-enabled device introduction by type of external collaboration partner |
Reading fidelity
high
Study strength
medium
|
not reported
|
| Firms that introduce AI-enabled medical devices have higher labour productivity. Firm Productivity | positive | Firm labour productivity |
Reading fidelity
high
Study strength
medium
|
not reported
|
| The positive labour-productivity effect of AI device introduction accumulates across successive device introductions. Firm Productivity | positive | Firm labour productivity as a function of cumulative AI device introductions |
Reading fidelity
high
Study strength
medium
|
not reported
|
| AI device introduction has a positive but weaker effect on profit margins, and the effect does not survive all model specifications. Firm Revenue | mixed | Firm profit margins |
Reading fidelity
high
Study strength
low
|
not reported
|
| The authors interpret the persistence of productivity gains alongside weaker margin gains as consistent with competitive entry eroding firms' pricing power as AI devices diffuse. Market Structure | negative | Profit margins and pricing power during AI device diffusion |
Reading fidelity
high
Study strength
speculative
|
not reported
|
| The FDA list used by the study contained 1,247 AI-enabled medical devices as of December 2025, approximately 94% of which entered through the 510(k) pathway. Adoption Rate | positive | Number and regulatory pathway distribution of AI-enabled medical devices |
Reading fidelity
high
Study strength
high
|
n=1247
approximately 94%
|
| Radiology, Cardiovascular, and Neurology account for more than 90% of FDA-cleared AI devices in the study's specialty-filtered sample. Adoption Rate | positive | Distribution of AI device approvals across medical specialties |
Reading fidelity
high
Study strength
high
|
n=1128
over 90%
|
| Within the three focal specialties, the FDA database contained 30,886 devices filed by 7,290 distinct applicants. Adoption Rate | positive | Number of medical devices and applicant firms in the focal specialties |
Reading fidelity
high
Study strength
high
|
n=7290
30,886 devices filed by 7,290 applicants
|
| The Orbis matching procedure produced 5,652 unique firm identifiers, corresponding to a 78% match rate. Organizational Efficiency | positive | Coverage of matched firm financial records |
Reading fidelity
high
Study strength
high
|
n=7290
78%
|