PaperlayerView exact revision on arXiv

Source-linked AI summary

Quantifying the impact of clinical-academic collaborations

Mohamad Zeina, Nick McNally, Karl S. Peggs, Parashkev Nachev

arXiv:2609.17051v1cs.DLcs.SI

TL;DR

The impact of clinical-academic collaborations, including specific institutional pairings, has not been robustly quantified. This paper applies an open-data graphical and bibliometric framework to compare academic networks; it finds that UCLH/UCL’s most underserved partners gain the most from collaboration.

  • Problem

    The impact of clinical-academic collaborations, including specific institutional pairings, has not been robustly quantified.

    Source: Introduction · p. 3

  • Method

    The paper applies an open-data graphical and bibliometric framework to identify academic networks and compare their bibliometric impact.

    Source: Discussion · p. 14

  • Results

    UCLH/UCL’s most underserved partners gain the most from collaboration.

    Source: Discussion · p. 13

  • Takeaways & Limitations

    The framework supports faithful comparison between academic networks and identification of attributes that predict bibliometric impact.

    Source: Discussion · p. 14

    Estimating how infrastructure relates to collaborative benefit is not straightforward and requires explicit modelling of many potentially material effects.

    Source: Discussion · p. 14

Abstract

from arXiv · show

Academic collaboration is of self-evident value but requires a quantitative representation to be optimally guided by policy. No established methodological approach to such representation exists. Here we introduce a general framework of graphical and bibliometric analysis of open data for the task of quantifying the impact of academic networks, with NIHR Biomedical Research Centre (BRC) clinical-academic partnerships in England as the prototype. We define publication-level identities for the 20 English BRCs based on the conjunction of authors from each BRC's partner institutions. Drawing on bibliometric and administrative records, we characterise the graphical properties of each network, estimate what the university adds to the hospital's papers, what the partnership adds to the papers of relatively infrastructure-poor collaborating institutions, and how that gain depends on existing infrastructure. We apply our framework to the NIHR UCLH/UCL BRC as an exemplar. UCLH/UCL authored 20,985 network papers from April 2007, in collaboration with 9,868 distinct external partners over the whole record, forming the most central node of the graph of networks across England. University co-authored papers exhibited 1.6 times the field-weighted citation impact (FWCI) of hospital-only papers, and were 2.1 times as likely to be cited by a patent. Across 60 of the exemplar's most partnered with UK healthcare organisations, the benefit rose from 1.8 times where local NIHR infrastructure activity was densest to 3.4 times where it was sparsest, while impact without the exemplar varied little. Academic networks can be robustly identified from open data, enabling comparative analysis of collaborative impact. Applied to NIHR BRCs, the approach enables quantification of the impact across networks and reveals that benefit is most pronounced where infrastructure is least developed.

Introduction

Clinical-academic partnerships need principled quantitative assessment, but their specific impact has not been robustly quantified. The paper proposes an open-data framework to define and compare these networks and their published outputs.

Introduction

  • The NIHR established BRCs in 2007 as local clinical-academic translational research networks, later expanding to 20 and broadening collaboration toward institutions with less infrastructure.This expansion reflects an aim to distribute UK research more widely.Source: Introduction · p. 3
  • Collaborative papers exceed individual papers in citation rates, but neither the existence nor the specific pairing of clinical-academic institutional collaborations has been robustly quantified.The paper identifies bibliographic opacity and limited public availability of comparators faithful across cross-network analysis as obstacles.Source: Introduction · p. 3
  • The framework identifies each of the 20 English BRCs from publication metadata, assigning papers to a network when authors represent both its hospital and university partners.This public, objective criterion supports consistent comparisons of network attributes and outputs across sites.Source: Introduction · p. 3
  • The study asks whether universities add value to hospital papers, whether networks benefit external healthcare collaborators lacking NIHR infrastructure, and whether that benefit varies with existing infrastructure.The third question concerns the potential value of extending collaborative infrastructure.Source: Introduction · p. 3
  • The framework is applied to all 20 networks, with the NIHR UCLH/UCL BRC serving as a worked exemplar for examining the magnitude, nature, and variability of collaborative effects.The stated objective is to characterise partnerships quantitatively using open data.Source: Introduction · p. 3

Methods · Data sources · Establishing BRC bibliographic identity

The methods combine publication, attention, and infrastructure records with rules for identifying BRC-linked papers. Analyses use distinct time windows for output, network structure, and partner gains.

Data sources

  • The March 2026 OpenAlex snapshot contained 272.5 million works [11]; Altmetric’s February 2025 extract supplied patent, policy, and news mentions, and NIHR project data identified infrastructure.Source: Data sources · p. 4
  • FWCI was the primary scholarly-impact measure, normalized against works of the same OpenAlex primary field, document type, and publication year [14].The measure was OpenAlex’s supplied work-record fwci field.Source: Data sources · p. 4
  • Output and citation analyses included papers published from 1 April 2007; reach and network structure used the whole record, while partner gains used 2015–2024.The partner-gain window was chosen because Altmetric coverage was stable during that period.Source: Data sources · p. 4

Establishing BRC bibliographic identity

  • For each of 20 English BRCs, a paper was attributed non-exclusively to a network if authors named both its host NHS trust and lead university [2022–2027 cohort].The institutions were identified by OpenAlex institution identifiers.Source: Establishing BRC bibliographic identity · p. 4
  • Trust identities included their OpenAlex-recorded child institutions, while the university side used one identifier; five additional institution pairs were derived for comparison.The additional pairs included King’s College London–King’s College Hospital and Edinburgh–NHS Lothian, plus three US pairs.Source: Establishing BRC bibliographic identity · p. 4
  • Because OpenAlex often matched UCLH papers to UCL, author-written affiliation strings were classified with a greedy longest-match span classifier to distinguish university from hospital authors.A large language model independently checked 1,000 strings across four resolution mechanisms, with some cases adjudicated by hand.Source: Establishing BRC bibliographic identity · p. 4
  • Hospitals sharing a university were treated as separate networks, with their overlapping papers excluded from the host network but countable for multiple networks when independently eligible.Source: Establishing BRC bibliographic identity · pp. 4–5
  • The analysis used the 2022–2027 BRC cohort throughout; Guy’s and St Thomas’–King’s was not in that cohort, and Exeter joined only in December 2022.The authors note that these cohort changes have consequences for interpretation.Source: Establishing BRC bibliographic identity · p. 5

Network structure

The section characterizes the BRC networks by counting their papers and external institutions, then comparing networks in an England-wide graph. It also describes measures of network centrality and the overlap in external partners with the UCLH/UCL BRC.

Network structure

  • Each network was characterized by counting its unique papers and distinct external institutions using OpenAlex institution identifiers.Source: Network structure · p. 5
  • The England-wide graph contained 22 network nodes, with edges weighted by the number of papers shared by each pair of networks.The nodes represented the 20 BRC networks plus King’s and Edinburgh.Source: Network structure · p. 5
  • Eigenvector centrality was computed on the weighted graph, while betweenness centrality was computed on the unweighted graph.Source: Network structure · p. 5
  • For each network, the analysis also measured the share of its external partners, excluding core university and hospital identifiers, that the UCLH/UCL BRC reached.Source: Network structure · p. 5

The value of clinical-academic partnership

The analysis compared hospital-only and university-coauthored partner papers across the 20 networks, modelling citation and attention outcomes while accounting for publication year, field and author count.

The value of clinical-academic partnership

  • Partner papers were classified as hospital-only when they lacked a university-affiliated author, or with-university when they included at least one.The same division was used to quantify the university’s involvement in hospital research.Source: The value of clinical-academic… · p. 5
  • Outcomes comprised log(1 + citations), FWCI, and indicators of any patent citation, policy-document citation or news mention.Source: The value of clinical-academic… · p. 5
  • Ordinary least squares models included group, publication-year and primary-field fixed effects, log(1 + author count), and HC3 robust standard errors.Binary outcomes used linear probability models, so coefficients represent percentage-point differences; models were fitted for all 20 networks, alongside unadjusted ratios of means and incidence.Source: The value of clinical-academic… · p. 5

Comparing sites served vs unserved by NIHR early-phase infrastructure

This section defines which places count as served by NIHR early-phase infrastructure and how partnerships with organisations in unserved places are identified and assessed.

Comparing sites served vs unserved by NIHR early-phase infrastructure

  • The study identified 114 named NIHR infrastructure centres and classified BRCs, Clinical Research Facilities, HealthTech Research Centres, Experimental Cancer Medicine Centres and BioResource centres as early-phase infrastructure.Applied Research Collaborations, Health Protection Research Units and patient-safety centres were excluded as applied or delivery schemes; two early-phase schemes were absent from the dataset.Source: Comparing sites served vs unserved by… · p. 5
  • A city hosting at least one early-phase centre counted as served; the study used a 27-city set, including Guildford and Preston.The set comprised 25 cities identified from the dataset and official lists, plus Guildford and Preston, whose facilities appeared in later NIHR open data.Source: Comparing sites served vs unserved by… · p. 6
  • Scotland, Wales and Northern Ireland were treated as unserved; sensitivity analyses used an England-only count, alternative served sets and a 50-institution cap on consortium papers.The alternative served sets were conservative and BRC-city-only.Source: Comparing sites served vs unserved by… · p. 6
  • For each network, the study counted distinct UK partner institutions in unserved sites and the NHS or healthcare subset; an established partnership required five joint papers and a founding paper.The study also counted partnerships active in 2021–2024, defined as having at least one joint paper in that period.Source: Comparing sites served vs unserved by… · p. 6
  • The study measured partner distance and, for partnerships whose first joint paper appeared in 2007–2015, the share still active in 2021–2024.Source: Comparing sites served vs unserved by… · p. 6

Collaborative benefit to underserved organisations

The section describes methods for comparing healthcare organisations’ paper impact with and without collaboration with a university-hospital network, including a separate exemplar-focused analysis and a served-city comparison group.

Collaborative benefit to underserved organisations

  • The primary sample covered 40 UK healthcare organisations and 86,983 papers from 2015–2024, selecting organisations with at least 50 co-authorships with any of 20 university-hospital pairs, independently of UCLH/UCL BRC membership.The 20 pairs comprised 19 non-UCL networks plus King’s College London and King’s College Hospital.Source: Collaborative benefit to underserved… · p. 6
  • Papers were grouped by collaboration with the exemplar, another of the 20 pairs, or neither; the pooled mean FWCI ratio compared exemplar-collaborative with non-collaborative papers using organisation-clustered bootstrap intervals.The analysis also reported medians and per-organisation ratios for organisations meeting minimum paper-count thresholds in each arm.Source: Collaborative benefit to underserved… · p. 6
  • The analysis also compared rates of reaching the sample’s top FWCI decile across the three paper groups, with organisation-clustered bootstrap intervals for rate ratios.The exemplar arm’s rate was compared with both the other-pair arm and the neither arm.Source: Collaborative benefit to underserved… · p. 6
  • A separate earlier sample of 14 organisations selected through at least 50 exemplar co-authorships was assessed on FWCI, raw citations, patent citation, and policy citation using unadjusted and adjusted estimates.Adjustment residualised outcomes for publication year, primary field, author count, and organisation code using five-fold cross-fitted gradient-boosted regression.Source: Collaborative benefit to underserved… · p. 6
  • The covariate adjustment was not a fixed-effects or doubly robust causal estimator, and the same analysis was repeated for 24 NHS trusts in served cities as a comparison group.Standard errors were clustered by organisation.Source: Collaborative benefit to underserved… · p. 6

Dependence of benefit on local infrastructure

The analysis tested whether collaboration benefit depends on partner organisations’ local research infrastructure, using 60 UK healthcare organisations that co-publish substantially with the exemplar.

Dependence of benefit on local infrastructure

  • The 60 organisations comprised 18 BRC host trusts, 27 other trusts in served cities and 15 organisations in unserved places.The analysis focused on whether benefit depends on the partner organisation’s local research infrastructure.Source: Dependence of benefit on local… · p. 7
  • Four infrastructure axes were assessed: nearby NIHR-supported project activity, distance to the nearest BRC, regional gross value added per head and deprivation.Gross value added and deprivation were assessed for the 41 English organisations only.Source: Dependence of benefit on local… · p. 7
  • Organisations were divided into tertiles per axis, and with-versus-without-exemplar mean FWCI ratios were compared using clustered bootstrap intervals and adjusted interaction and rank-correlation analyses.Bootstrap intervals and interaction estimates were applied to the two axes that showed a gradient; rank correlation related organisation-level ratios to scarcity.Source: Dependence of benefit on local… · p. 7

Statistical reporting and reproducibility

Analyses were applied consistently across all 20 networks, with confidence intervals and uncertainty estimates reported using specified methods. Network definitions, derived tables, and scripts were made available with table-level provenance information.

Statistical reporting and reproducibility

  • All 20 networks were analysed identically, with descriptive metrics in Supplementary Table S2 and within-hospital contrasts in Supplementary Table S3.Source: Statistical reporting and… · p. 7
  • 95% confidence intervals used organisation-clustered bootstrap for ratios and HC3 or cluster-robust standard errors for regression coefficients.Source: Statistical reporting and… · p. 7
  • Python analyses used pandas, statsmodels, scikit-learn, and NetworkX; network definitions, figure tables, and producing scripts are available in the project repository, whose table manifests record script, window, and checksum.Source: Statistical reporting and… · p. 7

Results

Across the 20 English BRC networks, UCLH/UCL had the highest weighted eigenvector centrality. Its university partnership was associated with higher publication impact, especially for healthcare organisations in places with the least local research infrastructure.

Global comparative analysis

  • UCLH/UCL had the highest weighted eigenvector centrality among the 20 networks (0.599), followed by GOSH (0.569) and Oxford (0.237).Network output and citations varied considerably; the lower recent-year citation counts reflect less time for citations to accrue, not declining impact.Source: Global comparative analysis · pp. 7, 9, Figure 1 · p. 8

The value of clinical-academic partnership

  • With-university papers had 1.57 times the hospital-only FWCI (3.54 vs. 2.25); adjusted for publication year, field, and team size, they received 68% more citations and had 0.85 higher FWCI.They also had a 2.3 percentage-point higher probability of patent inclusion and a 6.4 percentage-point higher probability of news coverage; policy-document inclusion was lower by 0.5 percentage points (incidence ratio 0.78).Source: The value of clinical-academic… · p. 9

Comparing sites served vs unserved by NIHR early-phase infrastructure

  • In sites without NIHR early-phase infrastructure, UCLH/UCL collaborated with 688 UK institutions—more than Oxford (664) or Cambridge (595)—including 418 NHS or healthcare organisations across 237 towns and cities.59% of UCLH/UCL’s UK healthcare partners are in unserved places.Source: Comparing sites served vs unserved by… · p. 9

Collaborative benefit to underserved organisations

  • Across 40 underserved healthcare organisations, papers with the exemplar had mean FWCI 8.02 vs 2.31 without it, a pooled ratio of means of 3.48 (95% CI 2.36 to 4.93).In the top FWCI decile, rates were 30.0% with the exemplar, 16.9% with another BRC pair, and 6.7% with neither; corresponding rate ratios were 1.78 and 4.51.Source: Collaborative benefit to underserved… · pp. 10–11
  • Among 14 unserved-place organisations, papers coauthored with the exemplar had 4.04 times mean FWCI (10.73 vs 2.66), with an adjusted difference of 3.69 (SE 1.63, p = 0.02).Source: Collaborative benefit to underserved… · p. 11, Figure 4 · p. 12

The benefit is largest where infrastructure support is scarcest

  • Across 60 partner organisations (20 per tertile), the exemplar’s with-versus-without FWCI ratio rose from 1.83 (95% interval 1.50 to 2.21) to 3.42 (2.35 to 4.93) as local NIHR-supported activity decreased.Without UCLH/UCL, mean FWCI was 2.86, 2.54 and 2.45 across decreasing-activity tertiles; the interaction was 1.06 per tertile (p = 0.004).Source: The benefit is largest where… · p. 11, Figure 4 · p. 12

Discussion

The discussion presents an open-data framework for comparing clinical-academic networks and finds that collaboration benefits vary across fields and partner infrastructure, with the largest gains among underserved partners. It cautions that the observational findings are not causal and identifies richer network and causal models as future work.

Discussion

  • The framework identifies networks from publication metadata using author and institutional identities, enabling transparent, equitable comparison despite imperfect identifiers.This introductory study used co-authorship involving at least one representative of constituent institutions; more elaborate membership weighting is possible.Source: Discussion · p. 13
  • The open-data framework can extend to richer graph models, uncertainty estimation, and principled counterfactual analysis to support prescriptive collaboration policy.The authors identify explicit modelling of plausibly material effects and more sophisticated causal designs as subjects for further investigation.Source: Discussion · pp. 13–14
  • Across NIHR BRCs, academic-partner involvement increased attention over hospital-only research, with benefits varying across fields and not explained by network size alone.The analysis covers scholarly, enterprise, government, and public attention; the specific drivers of variation require further study.Source: Discussion · p. 13
  • Among underserved partners, collaboration benefits were substantial across endpoints and, after adjustment, significant for field-weighted impact and policy citation; gains rose as partner NIHR support fell.The exemplar’s most underserved partners gained the most.Source: Discussion · p. 13
  • The inverse gradient suggests collaboration engages more powerfully where need is greatest, and that how effective hubs are built and operated may matter as much as their scale and span.The discussion proposes enhancing effective hubs’ infrastructure or replicating their collaborative policies as approaches that may benefit underserved nodes.Source: Discussion · pp. 13–14
  • Because co-authorship is non-random, the analysis supports strong associations rather than causal effects, despite design features that address some selection concerns.The binary, city-level served/unserved classification is an approximation, and the gradient estimates come from the exemplar’s 60 UK healthcare co-author organisations.Source: Discussion · p. 14

Declarations

The authors disclosed institutional affiliations and funding, identified available data and code with a person-level data limitation, and stated that ethics approval was not required.

Funding

  • The work was supported by the NIHR UCLH Biomedical Research Centre, and the views expressed are those of the authors, not necessarily NIHR or Department of Health views.Source: Funding · p. 14

Competing interests

  • All authors are employed by UCL and/or UCLH and affiliated with the NIHR UCLH BRC, one of the 20 centres analysed and the exemplar reported in full; no other competing interests were declared.Source: Competing interests · p. 15

Data and code availability

  • OpenAlex and NIHR infrastructure project data are openly available, analysis materials and scripts are on GitHub, and person-level data are not included.Altmetric data were used under licence.Source: Data and code availability · p. 15

Ethics

  • The study used publicly available institutional bibliographic metadata, involved no human participants, and did not require ethical approval.Source: Ethics · p. 15
Loading 2609.17051v1…