Referral uptake after diabetic retinopathy screening with artificial intelligence-assisted care pathways: a systematic review and meta-analysis

Rahmati et al.'s meta-analysis did not explore mechanisms underpinning differences in referral uptake.

“A recent meta-analysis of four studies13–15,17 by Rahmati et al.23 showed positive effects of AI-assisted DR screening on referral uptake, but did not explore mechanisms underpinning these differences.”

• cites [23] Rahmati, M. et al. Artificial Intelligence improves follow-up appointment uptake for diabetic retinal assessment: a systematic review and meta-analysis. Eye 39, 2398–2406 (2025). sourceDisputed

  The Discussion explains that AI can provide real-time grading and inform patients of a need for referral sooner than human graders. The Conclusions explicitly suggest that instant results, rather than delayed communication of human-grader results, most likely explain higher follow-up uptake. The paper therefore does discuss a possible mechanism, although it does not establish one experimentally. (LLM, conf=0.84)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured

There was marked variation in timeframes for referral among included studies, ranging from 2 weeks for Li et al. in China to 12 months for Liu et al. in the USA.

“There was marked variation for timeframes among included studies ranging from 2 weeks for Li et al. in China18 up to 12 months for Liu et al. in the USA17.”

• cites [18] Li, J. et al. Integrated image-based deep learning and language models for primary diabetes care. Nat. Med. 30, 2886–2896 (2024). sourceSupported
• cites [17] Liu, J. et al. Diabetic retinopathy screening with automated retinal image analysis in a primary care setting improves adherence to ophthalmic care. Ophthalmol. Retin. 5, 71–77 (2021). sourceDisputed

  The Methods (page not provided) specify referral appointments within 1 month for vision-threatening disease and within 3 months for referable disease or inconclusive results—not 12 months. The 12-month period in the Methods and Results is a window for measuring attendance after referral, not a recommended referral timeframe. This verdict concerns only the timeframe attributed to Liu et al.; it does not assess the comparison with Li et al. (LLM, conf=0.86)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured

Two studies that replaced routine referrals with targeted, AI-assisted referrals showed greater effect sizes than studies with a manual grading comparator.

“Firstly, comparator types whereby two studies replaced routine referrals with targeted, AI-assisted referrals13,17 showed greater effect sizes than those studies with a manual grading comparator14–16,18.”

• cites [13] Wolf, R. M. et al. Autonomous artificial intelligence increases screening and follow-up for diabetic retinopathy in youth: the ACCESS randomized control trial. Nat. Commun. 15, 421 (2024). sourceSupported
• cites [17] Liu, J. et al. Diabetic retinopathy screening with automated retinal image analysis in a primary care setting improves adherence to ophthalmic care. Ophthalmol. Retin. 5, 71–77 (2021). sourceSupported
• cites [14] Mathenge, W. et al. Impact of artificial intelligence assessment of diabetic retinopathy on referral service uptake in a low-resource setting: the RAIDERS randomized trial. Ophthalmol. Sci. 2, 100168 (2022). sourceDisputed

  The Methods (Randomization and Masking; page unavailable) say AI determined referral eligibility in both arms. This study compared immediate AI feedback with delayed advice after human grading, not targeted AI-assisted referrals with routine referrals. The cross-study effect-size comparison is the citing authors’ analysis, not a finding attributed to this source. (LLM, conf=0.88)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured
• cites [15] Dow, E. R. et al. Artificial intelligence improves patient follow-up in a diabetic retinopathy screening program. Clin. Ophthalmol. 17, 3323–3330 (2023). sourceDisputed

  The Methods (Program Overview; page unavailable) say the program did not alter established referral-placement processes: patients were referred to ophthalmology when indicated under both AI and human workflows. This contradicts the characterization of this study as replacing routine referrals with targeted, AI-assisted referrals. The Results report greater follow-up under the AI workflow, but do not make the citing authors’ cross-study effect-size comparison. (LLM, conf=0.85)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured
• cites [16] Chotcomwongse, P. et al. Transforming non-digital, clinical workflows to detect and track vision-threatening diabetic retinopathy via a digital platform integrating artificial intelligence: implementation research. Ophthalmol. Ther. 14, 447–460 (2025). sourcePartly supported

  The Results report AI-based referrals and more referral-positive and true-positive cases in the AI group. But this study compared AI screening with manual screening on alternate weeks, not a replacement of routine referrals as distinct from a manual-grading comparator. The cross-study comparison of effect sizes is the citing authors’ analysis, not a finding of this source. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [18] Li, J. et al. Integrated image-based deep learning and language models for primary diabetes care. Nat. Med. 30, 2886–2896 (2024). sourcePartly supported

  The Results (Prospective real-world study, experiment 2d; page unavailable) report higher ophthalmologist-attendance rates after referral with PCP+DeepDR-LLM than with unassisted PCPs. The paper does not show that routine referrals were replaced by targeted AI-assisted referrals. The comparison of effect sizes across the cited studies is the citing authors’ analysis, not a finding reported by this source. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

In settings with pre-existing manual grading workflows, AI-assisted DR screening provided incremental optimisation rather than transformation.

“Conversely, in settings with pre-existing manual grading workflows, AI-assisted DR screening provided incremental optimisation rather than transformation”

• cites [14] Mathenge, W. et al. Impact of artificial intelligence assessment of diabetic retinopathy on referral service uptake in a low-resource setting: the RAIDERS randomized trial. Ophthalmol. Sci. 2, 100168 (2022). sourcePartly supported

  Methods (Randomization and Masking; page not specified) compares immediate AI feedback with referral advice after human grading, and Results (page not specified) reports higher referral adherence with AI feedback (51.5% versus 39.6%). The paper does not establish that manual grading was a pre-existing workflow across the settings invoked by the claim, or characterize AI’s effect as incremental optimisation rather than transformation. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [15] Dow, E. R. et al. Artificial intelligence improves patient follow-up in a diabetic retinopathy screening program. Clin. Ophthalmol. 17, 3323–3330 (2023). sourcePartly supported

  The Methods (Program Overview; page unavailable) describe an existing human-graded workflow and a faster AI workflow, supporting the comparison. But the Results (page unavailable) report approximately three times higher ophthalmology follow-up under AI; the paper does not characterize this as merely incremental optimisation or assess whether it was transformative. (LLM, conf=0.83)

  Experimental — accuracy not yet measured
• cites [16] Chotcomwongse, P. et al. Transforming non-digital, clinical workflows to detect and track vision-threatening diabetic retinopathy via a digital platform integrating artificial intelligence: implementation research. Ophthalmol. Ther. 14, 447–460 (2025). sourceDisputed

  The Key Summary Points describe digitising existing non-digital workflows that had used manual grading as a transformation, not merely incremental optimisation. The Conclusions likewise say that digitising the conventional workflow through the platform is warranted, while noting possible performance improvements. The paper does not isolate AI's effect from the platform's. (LLM, conf=0.84)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured

Economic modelling studies of artificial intelligence diabetic retinopathy screening are limited by a lack of real-world data to inform parameter selection including transition probabilities, utility values, regional prevalence and uptake of screening and follow-up.

“However, these economic modelling studies are limited by a lack of real-world data to inform parameter selection including transition probabilities, utility values, regional prevalence and uptake of screening and follow-up10.”

• cites [10] Leigh, J., Drinkwater, J., Turner, A. & Schroeder, E.-A. Health Economic considerations for the implementation of artificial intelligence-enabled diabetic retinopathy screening: a review. Clin. Exp. Ophthalmol. 54, 144–161 (2026). sourcePartly supported

  The Results (Types of Economic Analyses and Modelling) report that transition probabilities often came from other countries, and the quality assessment identifies problems with utility-value sources and the transferability of disease-prevalence assumptions. Results (Implementation and Adherence Considerations) explicitly describe limited real-world data on follow-up adherence. The review does not clearly identify a lack of real-world data on initial screening uptake, so that item in the claim is not supported. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

Artificial intelligence is an accurate screening tool for diabetic retinopathy.

“Artificial intelligence (AI) is an accurate screening tool for diabetic retinopathy (DR), the leading cause of blindness among working-aged adults.”

• cites [5] Bhaskaranand, M. et al. The value of automated diabetic retinopathy screening with the eyeart system: a study of more than 100,000 consecutive encounters from people with diabetes. Diab. Technol. Ther. 21, 635 (2019). sourceSupported
• cites [6] Ruamviboonsuk, P. et al. Real-time diabetic retinopathy screening by deep learning in a multisite national screening programme: a prospective interventional cohort study. Lancet Digit. Heal. 4, e235–e244 (2022). sourceNot checked — no full text of the source
• cites [7] van der Heijden, A. A. et al. Validation of automated screening for referable diabetic retinopathy with the IDx-DR device in the Hoorn diabetes care system. Acta Ophthalmol. 96, 63–68 (2018). sourceSupported
• cites [8] Whitestone, N. et al. Feasibility and acceptance of artificial intelligence-based diabetic retinopathy screening in Rwanda. Br. J. Ophthalmol. 108, 840–845 (2024). sourceSupported

Referral uptake for those with referable diabetic retinopathy is sub-optimal.

“Indeed, beyond challenges of access and attendance at routine DR screening, referral uptake for those with referable DR is sub-optimal11,12 resulting in unmanaged patients at risk of blindness.”

• cites [11] Rahmati, M. et al. Factors affecting global adherence for the uptake of diabetic retinopathy screening: a systematic review and meta-analysis. Am. J. Ophthalmol. 268, 94–107 (2024). sourceCould not verify

  The Eligibility Criteria and §3.2 measure attendance at routine diabetic retinopathy screening among people with diabetes, not uptake of referrals after referable retinopathy is detected. Searches for referable DR and referral follow-up found no passage reporting referral uptake. (LLM, conf=0.75)

  Experimental — accuracy not yet measured
• cites [12] Bresnick, G. et al. Adherence to ophthalmology referral, treatment and follow-up after diabetic retinopathy screening in the primary care setting. BMJ Open Diab. Res. Care 8, e001154 (2020). sourceSupported

A recent meta-analysis of four studies by Rahmati et al. showed positive effects of artificial intelligence-assisted diabetic retinopathy screening on referral uptake.

“A recent meta-analysis of four studies13–15,17 by Rahmati et al.23 showed positive effects of AI-assisted DR screening on referral uptake, but did not explore mechanisms underpinning these differences.”

• cites [23] Rahmati, M. et al. Artificial Intelligence improves follow-up appointment uptake for diabetic retinal assessment: a systematic review and meta-analysis. Eye 39, 2398–2406 (2025). sourceSupported

Sub-optimal referral uptake results in unmanaged patients at risk of blindness.

“Indeed, beyond challenges of access and attendance at routine DR screening, referral uptake for those with referable DR is sub-optimal11,12 resulting in unmanaged patients at risk of blindness.”

• cites [11] Rahmati, M. et al. Factors affecting global adherence for the uptake of diabetic retinopathy screening: a systematic review and meta-analysis. Am. J. Ophthalmol. 268, 94–107 (2024). sourcePartly supported

  The Discussion (page unavailable) reports suboptimal uptake of routine diabetic retinopathy screening, and the Introduction (page unavailable) says prompt treatment can reduce sight-threatening disease and visual impairment. Neither passage establishes that uptake of referrals among patients with referable retinopathy is suboptimal or that missed referrals leave those patients unmanaged. (LLM, conf=0.78)

  Experimental — accuracy not yet measured
• cites [12] Bresnick, G. et al. Adherence to ophthalmology referral, treatment and follow-up after diabetic retinopathy screening in the primary care setting. BMJ Open Diab. Res. Care 8, e001154 (2020). sourceSupported

Both RCTs by Mathenge et al. and Wolf et al. were deemed low risk of bias as per RoB-2.

“Both RCTs by Mathenge et al.14 and Wolf et al.13 were deemed low risk of bias as per the RoB-2.”

• cites [14] Mathenge, W. et al. Impact of artificial intelligence assessment of diabetic retinopathy on referral service uptake in a low-resource setting: the RAIDERS randomized trial. Ophthalmol. Sci. 2, 100168 (2022). sourceSupported
• cites [13] Wolf, R. M. et al. Autonomous artificial intelligence increases screening and follow-up for diabetic retinopathy in youth: the ACCESS randomized control trial. Nat. Commun. 15, 421 (2024). sourceSupported

Real-world implementation of diabetic retinopathy screening with artificial intelligence-assisted care pathways has accelerated globally since 2020.

“Real-world implementation of DR screening with AI-assisted care pathways has accelerated globally since 20208,13–22, yet evidence on clinical and health system outcomes such as referral uptake remains limited.”

• cites [8] Whitestone, N. et al. Feasibility and acceptance of artificial intelligence-based diabetic retinopathy screening in Rwanda. Br. J. Ophthalmol. 108, 840–845 (2024). sourcePartly supported

  The Abstract describes AI-based diabetic retinopathy screening at four clinics in Rwanda from March to June 2021, providing a real-world example after 2020. It does not establish that implementation accelerated globally or compare implementation rates over time. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [13] Wolf, R. M. et al. Autonomous artificial intelligence increases screening and follow-up for diabetic retinopathy in youth: the ACCESS randomized control trial. Nat. Commun. 15, 421 (2024). sourcePartly supported

  The Introduction (page unavailable) describes implementation of autonomous AI screening in a multidisciplinary diabetes clinic, and the Discussion (page unavailable) reports results from the ACCESS trial. These passages support real-world use, but they do not establish that implementation accelerated globally since 2020. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [14] Mathenge, W. et al. Impact of artificial intelligence assessment of diabetic retinopathy on referral service uptake in a low-resource setting: the RAIDERS randomized trial. Ophthalmol. Sci. 2, 100168 (2022). sourcePartly supported

  The Methods report AI-assisted diabetic retinopathy screening at four clinics in Rwanda in 2021, supporting this paper as an example of real-world implementation after 2020. The paper does not establish that implementation accelerated globally; its introduction says evidence from real-life programs was limited and that Africa lagged in clinical deployment. Page numbers were not available in our copy. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [15] Dow, E. R. et al. Artificial intelligence improves patient follow-up in a diabetic retinopathy screening program. Clin. Ophthalmol. 17, 3323–3330 (2023). sourcePartly supported

  The Methods (Program Overview; page not specified) describe a real-world diabetic retinopathy screening program in the San Francisco Bay Area that adopted an AI workflow after its first 18 months, which began in September 2019. The paper supports an example of implementation after 2020, but does not establish that implementation accelerated globally since 2020. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [16] Chotcomwongse, P. et al. Transforming non-digital, clinical workflows to detect and track vision-threatening diabetic retinopathy via a digital platform integrating artificial intelligence: implementation research. Ophthalmol. Ther. 14, 447–460 (2025). sourcePartly supported

  The Methods and Introduction describe a real-world AI-assisted diabetic retinopathy screening pathway, and the Discussion summarizes studies in several countries. The paper does not establish that implementation accelerated globally since 2020; the listed studies do not by themselves demonstrate that trend. Page numbers are not available in our copy. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [17] Liu, J. et al. Diabetic retinopathy screening with automated retinal image analysis in a primary care setting improves adherence to ophthalmic care. Ophthalmol. Retin. 5, 71–77 (2021). sourcePartly supported

  The Methods describe AI-assisted screening in one St. Louis primary care clinic, with patients recruited from January to August 2018. This supports an example of real-world implementation, but not the claim that implementation accelerated globally since 2020. (LLM, conf=0.88)

  Experimental — accuracy not yet measured
• cites [18] Li, J. et al. Integrated image-based deep learning and language models for primary diabetes care. Nat. Med. 30, 2886–2896 (2024). sourcePartly supported

  The Methods (Real-world prospective study; page unavailable) describe deployment of an AI-assisted diabetes care and DR screening system at one Chinese primary-care site in 2023. The Methods (Data acquisition and diagnosis criteria; page unavailable) also identify real-world DR screening programs in China. These examples support real-world implementation, but the paper does not establish that implementation accelerated globally since 2020. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [19] Heuer, R. et al. AI diabetic retinopathy screening in a primary care setting in rural maine. J. Gen. Intern. Med. 40, 2465–2467 (2025). sourcePartly supported

  The Methods describe a real-world AI-assisted screening and referral pathway at one rural Maine clinic from 2021 to 2023. The paper does not establish that implementation accelerated globally since 2020; its stated scope is a single healthcare system in one state. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [20] Chauhan, A. et al. Barriers and determinants of referral adherence in AI-enabled diabetic retinopathy screening for older adults in northern india during the COVID-19 pandemic: mixed methods pilot study. JMIR Formative Res. 9, e67047 (2025). sourcePartly supported

  The Methods (Hardware and Software for DRS; page unavailable) document real-world AI-assisted screening at two sites in Punjab. The Introduction says the pandemic emphasized the need for AI-enabled screening, but neither passage establishes that implementation accelerated globally since 2020. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [21] Scheetz, J. et al. Real-world artificial intelligence-based opportunistic screening for diabetic retinopathy in endocrinology and indigenous healthcare settings in Australia. Sci. Rep. 11, 15808 (2021). sourcePartly supported

  The Introduction (page not supplied) describes international real-world use of AI-based diabetic retinopathy screening as “early and emerging,” and the study documents implementation in Australian clinics. It does not establish that implementation accelerated globally or identify 2020 as the start of that acceleration. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

Studies comparing AI-assisted screening to existing manual grading workflows showed modest gains ranging from +11.8 to +23.5%.

“Conversely, studies comparing AI-assisted screening to existing manual grading workflows showed modest gains ranging from +11.8 to +23.5%14–16,18.”

• cites [14] Mathenge, W. et al. Impact of artificial intelligence assessment of diabetic retinopathy on referral service uptake in a low-resource setting: the RAIDERS randomized trial. Ophthalmol. Sci. 2, 100168 (2022). sourcePartly supported

  The Methods (Randomization and Masking; page unavailable) compare immediate AI feedback with referral advice after human grading. The Results (page unavailable) report referral adherence of 51.5% versus 39.6%—about 11.9 percentage points higher—but describe this as a 30.1% relative increase. This paper therefore supports a gain near the stated lower bound, not the precise +11.8% to +23.5% range attributed to the cited studies. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [15] Dow, E. R. et al. Artificial intelligence improves patient follow-up in a diabetic retinopathy screening program. Clin. Ophthalmol. 17, 3323–3330 (2023). sourceSupported
• cites [16] Chotcomwongse, P. et al. Transforming non-digital, clinical workflows to detect and track vision-threatening diabetic retinopathy via a digital platform integrating artificial intelligence: implementation research. Ophthalmol. Ther. 14, 447–460 (2025). sourcePartly supported

  Results, Table 2 reports referral adherence among true positives of 89.1% with AI screening versus 77.3% with manual screening—a difference of 11.8 percentage points—but says the difference was not statistically significant. The paper does not report a 23.5-point gain, and the Discussion cautions that the different workflows may not be directly comparable. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [18] Li, J. et al. Integrated image-based deep learning and language models for primary diabetes care. Nat. Med. 30, 2886–2896 (2024). sourcePartly supported

  The Results section (‘DeepDR-Transformer as an assistive tool,’ experiment 2c; page unavailable) compares grading with and without AI assistance and reports improved sensitivity and specificity. It does not report gains ranging from +11.8% to +23.5%; searches for both figures found no matching result in our copy. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

All studies consistently demonstrate improved referral uptake after AI-assisted DR screening.

“While all studies consistently demonstrate improved referral uptake after AI-assisted DR screening13–18, magnitude of benefit varies substantially from moderate improvements ( + 11.8%)16 to large effects ( + 42.0%)13.”

• cites [13] Wolf, R. M. et al. Autonomous artificial intelligence increases screening and follow-up for diabetic retinopathy in youth: the ACCESS randomized control trial. Nat. Commun. 15, 421 (2024). sourceSupported
• cites [14] Mathenge, W. et al. Impact of artificial intelligence assessment of diabetic retinopathy on referral service uptake in a low-resource setting: the RAIDERS randomized trial. Ophthalmol. Sci. 2, 100168 (2022). sourceSupported
• cites [15] Dow, E. R. et al. Artificial intelligence improves patient follow-up in a diabetic retinopathy screening program. Clin. Ophthalmol. 17, 3323–3330 (2023). sourceSupported
• cites [16] Chotcomwongse, P. et al. Transforming non-digital, clinical workflows to detect and track vision-threatening diabetic retinopathy via a digital platform integrating artificial intelligence: implementation research. Ophthalmol. Ther. 14, 447–460 (2025). sourcePartly supported

  The Results (Table 2) show higher referral adherence with AI screening than manual screening: 89.1% versus 77.3%, a difference of 11.8 percentage points. The difference was not statistically significant (P = 0.158), so this study supports an observed increase but not the stronger assertion that it demonstrates improved uptake. This paper alone also cannot establish what all cited studies found. (LLM, conf=0.88)

  Experimental — accuracy not yet measured
• cites [17] Liu, J. et al. Diabetic retinopathy screening with automated retinal image analysis in a primary care setting improves adherence to ophthalmic care. Ophthalmol. Retin. 5, 71–77 (2021). sourceSupported
• cites [18] Li, J. et al. Integrated image-based deep learning and language models for primary diabetes care. Nat. Med. 30, 2886–2896 (2024). sourceSupported

Several studies included co-interventions alongside AI-assisted diabetic retinopathy screening with variable effectiveness.

“Several studies included co-interventions alongside AI-assisted DR screening with variable effectiveness13,16–18.”

• cites [13] Wolf, R. M. et al. Autonomous artificial intelligence increases screening and follow-up for diabetic retinopathy in youth: the ACCESS randomized control trial. Nat. Commun. 15, 421 (2024). sourceSupported
• cites [16] Chotcomwongse, P. et al. Transforming non-digital, clinical workflows to detect and track vision-threatening diabetic retinopathy via a digital platform integrating artificial intelligence: implementation research. Ophthalmol. Ther. 14, 447–460 (2025). sourceSupported
• cites [17] Liu, J. et al. Diabetic retinopathy screening with automated retinal image analysis in a primary care setting improves adherence to ophthalmic care. Ophthalmol. Retin. 5, 71–77 (2021). sourceSupported
• cites [18] Li, J. et al. Integrated image-based deep learning and language models for primary diabetes care. Nat. Med. 30, 2886–2896 (2024). sourcePartly supported

  The Results (experiment 2d) describe one study that combined AI-assisted diabetic retinopathy assessment with AI-enhanced primary-care recommendations, and report improved referral follow-through. The paper therefore supports an example of an added intervention alongside screening, but does not establish that several studies used such interventions or that their effectiveness varied. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

In higher-performing systems (baseline referral uptake 39.6-77.3%), the gains from AI-assisted screening compared to manual grading were modest.

“This was most pronounced in in higher-performing systems (baseline referral uptake 39.6–77.3%)14,16,18 which indicates diminishing returns in settings with already functional screening infrastructure.”

• cites [14] Mathenge, W. et al. Impact of artificial intelligence assessment of diabetic retinopathy on referral service uptake in a low-resource setting: the RAIDERS randomized trial. Ophthalmol. Sci. 2, 100168 (2022). sourcePartly supported

  The Results (page unavailable) report referral uptake of 39.6% with delayed human-grader feedback versus 51.5% with immediate AI-supported feedback, an 11.9-percentage-point gain. This supports the lower endpoint and permits a description of the absolute gain as modest, but the paper studies a low-resource setting; it does not characterize that setting as a higher-performing system or establish the stated 39.6–77.3% range. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [16] Chotcomwongse, P. et al. Transforming non-digital, clinical workflows to detect and track vision-threatening diabetic retinopathy via a digital platform integrating artificial intelligence: implementation research. Ophthalmol. Ther. 14, 447–460 (2025). sourceSupported
• cites [18] Li, J. et al. Integrated image-based deep learning and language models for primary diabetes care. Nat. Med. 30, 2886–2896 (2024). sourcePartly supported

  The Results (Prospective real-world study, experiment 2d; page unavailable) report referral follow-through of 58.44% without DeepDR-LLM and 77.78% with it. The unassisted figure falls within the cited baseline range, but this paper does not establish the range’s endpoints or characterize the gain as modest in higher-performing systems. (LLM, conf=0.79)

  Experimental — accuracy not yet measured

The magnitude of benefit from AI-assisted DR screening varies substantially from moderate improvements (+11.8%) to large effects (+42.0%).

“While all studies consistently demonstrate improved referral uptake after AI-assisted DR screening13–18, magnitude of benefit varies substantially from moderate improvements ( + 11.8%)16 to large effects ( + 42.0%)13.”

• cites [16] Chotcomwongse, P. et al. Transforming non-digital, clinical workflows to detect and track vision-threatening diabetic retinopathy via a digital platform integrating artificial intelligence: implementation research. Ophthalmol. Ther. 14, 447–460 (2025). sourceSupported
• cites [13] Wolf, R. M. et al. Autonomous artificial intelligence increases screening and follow-up for diabetic retinopathy in youth: the ACCESS randomized control trial. Nat. Commun. 15, 421 (2024). sourceSupported

Li et al.'s sequential cohort study (n=298) found 77.8% versus 58.4% referral uptake within two weeks for AI-assisted screening with treatment advice based on multimodal patient data and retinal images versus unassisted manual grading (+19.4%, p=0.001).

“Li et al.‘s sequential cohort study (n = 298) compared AI-assisted screening with treatment advice based upon multimodal patient data (medical history, physical examinations, laboratory tests) and retinal images, to unassisted manual grading, achieving 77.8% versus 58.4% referral uptake within two weeks (+19.4%, p = 0.001)18.”

• cites [18] Li, J. et al. Integrated image-based deep learning and language models for primary diabetes care. Nat. Med. 30, 2886–2896 (2024). sourcePartly supported

  Split into 2 atoms — 1 confirmed, 0 partially supported, 0 disputed, 0 misattributed, 0 unverifiable, 1 not checked. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

Li et al., despite allowing just a fortnight for appointments, reported referral uptake of 58.4% and 77.8% for comparator and intervention groups, respectively.

“For example, Li et al., despite allowing just a fortnight for appointments reported referral uptake of 58.4% and 77.8% for comparator and intervention groups, respectively18.”

• cites [18] Li, J. et al. Integrated image-based deep learning and language models for primary diabetes care. Nat. Med. 30, 2886–2896 (2024). sourceSupported

In Liu et al., there was a correlation between time and rates of uptake, increasing from 32.6% at 3 months to 55.4% at 12 months.

“In this study there was a correlation between time and rates of uptake, increasing from 32.6% at 3 months to 55.4% at 12 months17.”

• cites [17] Liu, J. et al. Diabetic retinopathy screening with automated retinal image analysis in a primary care setting improves adherence to ophthalmic care. Ophthalmol. Retin. 5, 71–77 (2021). sourceSupported

Other US reports show 5-51% referral uptake within recommended timeframes.

“These low baseline rates of referral uptake align with other US reports showing 5–51% referral uptake within recommended timeframes12, indicating substantial potential benefit in these settings.”

• cites [12] Bresnick, G. et al. Adherence to ophthalmology referral, treatment and follow-up after diabetic retinopathy screening in the primary care setting. BMJ Open Diab. Res. Care 8, e001154 (2020). sourcePartly supported

  The Results > Adherence rates section reports that 5% kept a first ophthalmology appointment within the recommended referral interval. It reports 48% and 56% within one year in two study periods, but one year was not the recommended interval for those patients. The cited paper therefore does not support describing the 51% figure as referral uptake within recommended timeframes. (LLM, conf=0.90)

  Experimental — accuracy not yet measured

Chotcomwongse et al.'s AI platform provided automated text message reminders to patients with referable diabetic retinopathy several days before their scheduled referral appointment.

“Chotcomwongse et al.’s AI platform provided automated text message reminders to patients with referable DR several days before their scheduled referral appointment16.”

• cites [16] Chotcomwongse, P. et al. Transforming non-digital, clinical workflows to detect and track vision-threatening diabetic retinopathy via a digital platform integrating artificial intelligence: implementation research. Ophthalmol. Ther. 14, 447–460 (2025). sourceSupported

Wolf et al. reported a +42.0% improvement in referral uptake in a diabetology clinic screening youth with diabetes.

“Wolf et al. reported a + 42.0% improvement in a diabetology clinic screening youth with diabetes13”

• cites [13] Wolf, R. M. et al. Autonomous artificial intelligence increases screening and follow-up for diabetic retinopathy in youth: the ACCESS randomized control trial. Nat. Commun. 15, 421 (2024). sourceSupported

Chotcomwongse et al.'s alternating-week study (n=304) found 89.1% versus 77.3% referral uptake for AI screening coupled with automated text reminders versus nurse-led manual grading (+11.8%, p=0.158, not statistically significant).

“Chotcomwongse et al.‘s alternating-week study design (n = 304) compared AI screening coupled with automated text reminders before scheduled follow-up to nurse-led manual grading, finding 89.1% versus 77.3% referral uptake (+11.8%, p = 0.158, not statistically significant)16.”

• cites [16] Chotcomwongse, P. et al. Transforming non-digital, clinical workflows to detect and track vision-threatening diabetic retinopathy via a digital platform integrating artificial intelligence: implementation research. Ophthalmol. Ther. 14, 447–460 (2025). sourcePartly supported

  The Methods (Study Design; page not provided) describe AI and manual screening on alternate weeks, and the Methods (Implementation and Variables; page not provided) describe automated phone reminders for AI-positive patients. Results Table 2 reports 89.1% versus 77.3% referral adherence among true positives (an 11.8-percentage-point difference; p=0.158). The stated n=304 is incorrect: Results Table 1 lists 708 AI-screened and 746 manually screened patients. (LLM, conf=0.92)

  Experimental — accuracy not yet measured

Liu et al. reported a +36.7% improvement in referral uptake in a primary care clinic screening adults with diabetes.

“whilst Liu et al. reported a + 36.7% improvement in a primary care clinic screening adults with diabetes17.”

• cites [17] Liu, J. et al. Diabetic retinopathy screening with automated retinal image analysis in a primary care setting improves adherence to ophthalmic care. Ophthalmol. Retin. 5, 71–77 (2021). sourceSupported

Mathenge et al.'s RAIDERS RCT (n=275) found 51.5% versus 39.6% referral uptake for immediate AI result acquisition with colour-coded report versus delayed manual grading (3-5 days) with SMS and phone calls (+11.9%, OR 1.62, 95% CI 1.00-2.61, p=0.048).

“Mathenge et al.‘s RAIDERS RCT (n = 275) uniquely isolated the effect of immediate versus delayed feedback by randomising patients to immediate AI result acquisition with a colour-coded report based upon severity, or delayed manual grading (3–5 days) with SMS and phone calls to flag results, finding 51.5% versus 39.6% referral uptake respectively ( + 11.9%, OR 1.62, 95% CI: 1.00–2.61, p = 0.048)14.”

• cites [14] Mathenge, W. et al. Impact of artificial intelligence assessment of diabetic retinopathy on referral service uptake in a low-resource setting: the RAIDERS randomized trial. Ophthalmol. Sci. 2, 100168 (2022). sourceSupported

Dow et al.'s retrospective study (n=396) showed 35.5% referral uptake with AI workflow versus 12.0% with manual grading when restricting analysis to university eye institute appointments (+23.5%, p=0.000004).

“Dow et al.‘s retrospective study (n = 396) showed 35.5% referral uptake with AI workflow versus 12.0% with manual grading when restricting analysis to university eye institute appointments (+23.5%, p = 0.000004)15.”

• cites [15] Dow, E. R. et al. Artificial intelligence improves patient follow-up in a diabetic retinopathy screening program. Clin. Ophthalmol. 17, 3323–3330 (2023). sourcePartly supported

  The Results report 35.5% (99/279) follow-up at the university eye institute under the AI Workflow versus 12.0% (14/117) under the Human Workflow; those denominators total 396, and the difference is 23.5 percentage points. The study describes its analyses as retrospective (Discussion). However, the Results give the pairwise comparison as p < 0.000004, not p = 0.000004. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

Facilitated scheduling addresses social opportunity barriers.

“Facilitated scheduling addresses social opportunity barriers16,17”

• cites [16] Chotcomwongse, P. et al. Transforming non-digital, clinical workflows to detect and track vision-threatening diabetic retinopathy via a digital platform integrating artificial intelligence: implementation research. Ophthalmol. Ther. 14, 447–460 (2025). sourcePartly supported

  The Discussion (page unavailable) describes facilitating appointments as a benefit of the digital platform. It does not identify social opportunity barriers or show that facilitated scheduling addresses them, so that part of the claim is not supported by this paper. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [17] Liu, J. et al. Diabetic retinopathy screening with automated retinal image analysis in a primary care setting improves adherence to ophthalmic care. Ophthalmol. Retin. 5, 71–77 (2021). sourcePartly supported

  The Methods (page unavailable) describe staff calling patients to schedule follow-up ophthalmology appointments, with repeat calls after unsuccessful contact. The Discussion reports improved attendance for the overall screening program, but the paper does not identify social opportunity barriers or establish that facilitated scheduling itself addressed them. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

Immediate AI results catalyse coordinated interventions across multiple behavioural domains.

“Immediate AI results catalyse coordinated interventions across multiple behavioural domains25 (Fig. 4).”

• cites [25] Michie, S., van Stralen, M. M. & West, R. The behaviour change wheel: a new method for characterising and designing behaviour change interventions. Implement. Sci. 6, 42 (2011). sourcePartly supported

  The Methods section (Establishing criteria of usefulness; page unavailable) says an intervention may target multiple components of the behaviour system, and the Results section (Systematic literature review of existing frameworks; page unavailable) says one intervention may perform several behaviour-change functions. These passages support the general idea of a multifaceted intervention, but they do not address immediate AI results or show that such results catalyse coordinated interventions. (LLM, conf=0.78)

  Experimental — accuracy not yet measured

Point-of-care counselling enhances psychological capability.

“point-of-care counselling enhances psychological capability13”

• cites [13] Wolf, R. M. et al. Autonomous artificial intelligence increases screening and follow-up for diabetic retinopathy in youth: the ACCESS randomized control trial. Nat. Commun. 15, 421 (2024). sourcePartly supported

  The Discussion (page unavailable) reports that patients given immediate point-of-care results and education had higher follow-through with eye care. It does not assess psychological capability or establish that counselling enhanced it; the measured outcome was follow-through. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

Mathenge et al. found that immediate versus delayed feedback after AI-assisted screening increased referral uptake by 11.9%.

“Mathenge et al.’s isolation of immediate versus delayed feedback after AI-assisted screening ( + 11.9% referral uptake)”

• cites [14] Mathenge, W. et al. Impact of artificial intelligence assessment of diabetic retinopathy on referral service uptake in a low-resource setting: the RAIDERS randomized trial. Ophthalmol. Sci. 2, 100168 (2022). sourceSupported

In settings where pre-existing workflows included routine referral strategies, AI-assisted DR screening enabled transformation of care pathways by replacing routine referrals with targeted, evidence-based referrals.

“In settings where pre-existing workflows included routine referral strategies13,17, AI-assisted DR screening enabled transformation of care pathways by replacing routine referrals with targeted, evidence-based referrals.”

• cites [13] Wolf, R. M. et al. Autonomous artificial intelligence increases screening and follow-up for diabetic retinopathy in youth: the ACCESS randomized control trial. Nat. Commun. 15, 421 (2024). sourceSupported
• cites [17] Liu, J. et al. Diabetic retinopathy screening with automated retinal image analysis in a primary care setting improves adherence to ophthalmic care. Ophthalmol. Retin. 5, 71–77 (2021). sourceSupported

Provision of personalised reports supports reflective motivation.

“provision of personalised reports supports reflective motivation18”

• cites [18] Li, J. et al. Integrated image-based deep learning and language models for primary diabetes care. Nat. Med. 30, 2886–2896 (2024). sourcePartly supported

  The Discussion says DeepDR-LLM generated recommendations that PCPs used to formulate management plans for individual patients, and the Results report better self-management in the assisted arm. Neither section assesses reflective motivation or establishes that personalised reports support that specific mechanism. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

Result timeliness alone modestly improves referral uptake through enhanced capability.

“demonstrates that result timeliness alone modestly improves uptake through enhanced capability14”

• cites [14] Mathenge, W. et al. Impact of artificial intelligence assessment of diabetic retinopathy on referral service uptake in a low-resource setting: the RAIDERS randomized trial. Ophthalmol. Sci. 2, 100168 (2022). sourcePartly supported

  The Results report higher referral uptake with immediate AI-supported feedback than with delayed communication (51.5% versus 39.6%). But the Methods describe differences beyond timeliness, and the Discussion suggests several possible contributors, including retinal images, counselling, and same-day visits. The paper does not establish that timeliness alone caused the improvement or that it worked through enhanced capability. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

Baseline health system performance influenced the marginal gains, with settings with higher baseline rates of uptake experiencing relatively lower effect size, in contrast to lower-performing systems with greater margins for improvement.

“Thirdly, baseline health system performance appeared to influence the marginal gains, following a law of diminishing returns whereby settings with higher baseline rates of uptake experienced relatively lower effect size14,16,18, in contrast to lower-performing systems with greater margins for improvement13,15,17.”

• cites [14] Mathenge, W. et al. Impact of artificial intelligence assessment of diabetic retinopathy on referral service uptake in a low-resource setting: the RAIDERS randomized trial. Ophthalmol. Sci. 2, 100168 (2022). sourcePartly supported

  The Results report 39.6% referral uptake in the control group versus 51.5% in the AI group, a 30.1% relative increase. The Discussion also compares uptake rates across settings, but the paper does not establish that baseline health-system performance influenced effect size or demonstrate a diminishing-returns pattern. (LLM, conf=0.78)

  Experimental — accuracy not yet measured
• cites [16] Chotcomwongse, P. et al. Transforming non-digital, clinical workflows to detect and track vision-threatening diabetic retinopathy via a digital platform integrating artificial intelligence: implementation research. Ophthalmol. Ther. 14, 447–460 (2025). sourceCould not verify

  The Results compare AI and manual screening referrals and referral adherence at one site; they do not report baseline uptake rates across health systems. The Discussion considers possible reasons for adherence but does not link higher baseline uptake to smaller gains. Searches for baseline screening uptake and diminishing returns found no passage addressing that relationship. (LLM, conf=0.76)

  Experimental — accuracy not yet measured
• cites [18] Li, J. et al. Integrated image-based deep learning and language models for primary diabetes care. Nat. Med. 30, 2886–2896 (2024). sourceCould not verify

  The Results section on the prospective real-world study reports higher referral follow-through with DeepDR-LLM than with unassisted primary care (77.78% versus 58.44%). The Methods describe baseline and follow-up comparisons within that study, but neither section relates differences in effect size to baseline uptake across health systems. Searches for diminishing returns and baseline health-system performance found no passage addressing that relationship; page numbers were unavailable. (LLM, conf=0.74)

  Experimental — accuracy not yet measured
• cites [13] Wolf, R. M. et al. Autonomous artificial intelligence increases screening and follow-up for diabetic retinopathy in youth: the ACCESS randomized control trial. Nat. Commun. 15, 421 (2024). sourcePartly supported

  The Results (Primary outcome; page unavailable) report 22% screening completion with standard care versus 100% with AI, a large gain from low usual-care uptake. The paper does not show that baseline performance determines effect size or compare gains across higher- and lower-performing health systems. (LLM, conf=0.78)

  Experimental — accuracy not yet measured
• cites [15] Dow, E. R. et al. Artificial intelligence improves patient follow-up in a diabetic retinopathy screening program. Clin. Ophthalmol. 17, 3323–3330 (2023). sourcePartly supported

  The Results report that follow-up rose from 12.0% under the Human Workflow to 35.5% under the AI Workflow, a three-fold increase. This supports the example of substantial improvement from a low starting rate, but the paper does not compare health systems or establish that baseline performance determines the size of gains. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [17] Liu, J. et al. Diabetic retinopathy screening with automated retinal image analysis in a primary care setting improves adherence to ophthalmic care. Ophthalmol. Retin. 5, 71–77 (2021). sourcePartly supported

  The Results report low historical adherence at the clinic (18.7%) and higher follow-up after automated screening; the Discussion describes the clinic as having a historically low screening rate. The paper studies one clinic and does not establish that higher-baseline health systems experience smaller gains or that baseline performance determines effect size across settings. Page numbers were not provided in our copy. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

AI's effectiveness varies by organisational context.

“AI’s effectiveness varies by organisational context, reflecting the NASSS Framework’s value proposition domain26.”

• cites [26] Greenhalgh, T. et al. Beyond adoption: a new framework for theorizing and evaluating nonadoption, abandonment, and challenges to the scale-up, spread, and sustainability of health and care technologies. J. Med. Internet Res. 19, e8775 (2017). sourcePartly supported

  In Results > Synthesis: The NASSS Framework (page not specified), the paper says organizational readiness influenced the uptake and scale-up of technology-supported programs. Its value proposition domain considers a technology’s efficacy, but the paper does not establish that AI’s effectiveness varies by organizational context; it treats organizational readiness separately under the organization domain. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

Dow et al. achieved a 23.5% improvement in referral uptake without documented co-interventions for behaviour change.

“Dow et al. achieved substantial improvement ( + 23.5%) without documented co-interventions for behaviour change15”

• cites [15] Dow, E. R. et al. Artificial intelligence improves patient follow-up in a diabetic retinopathy screening program. Clin. Ophthalmol. 17, 3323–3330 (2023). sourcePartly supported

  The Results report follow-up at the university eye institute of 35.5% under the AI workflow versus 12.0% under the human workflow—a 23.5-percentage-point difference, not a 23.5% relative improvement in referral uptake generally. The Methods (Program Overview) say the prescribed workflow differences concerned how and when results were communicated, supporting the absence of a documented separate behaviour-change intervention. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

This system-level transformation yielded large effect sizes (+36.7-42.0% absolute improvement in referral uptake).

“This system-level transformation yielded large effect sizes ( + 36.7–42.0% absolute improvement in referral uptake)13,17”

• cites [13] Wolf, R. M. et al. Autonomous artificial intelligence increases screening and follow-up for diabetic retinopathy in youth: the ACCESS randomized control trial. Nat. Commun. 15, 421 (2024). sourcePartly supported

  The Results (Secondary outcomes; page unavailable) report a 42-percentage-point difference in follow-through completion: 64% in the AI intervention arm versus 22% in the control arm. This supports the upper end of the claimed referral-uptake range, but our copy does not report a 36.7-percentage-point improvement. (LLM, conf=0.88)

  Experimental — accuracy not yet measured
• cites [17] Liu, J. et al. Diabetic retinopathy screening with automated retinal image analysis in a primary care setting improves adherence to ophthalmic care. Ophthalmol. Retin. 5, 71–77 (2021). sourcePartly supported

  The Abstract Results report 55.4% adherence to follow-up eye care at one year among referred patients, versus 18.7% in historical controls—an absolute difference of 36.7 percentage points. The paper does not report the 42.0-point end of the cited range. Its Discussion also cautions that selection bias limits a causal interpretation of the comparison. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

Workflow redesign alone may drive gains in referral uptake in some contexts, even with existing manual grading infrastructure such as asynchronous telehealth via remote grading centres.

“suggesting workflow redesign alone may drive gains in referral uptake in some contexts, even with existing manual grading infrastructure such as asynchronous telehealth via remote grading centres.”

• cites [15] Dow, E. R. et al. Artificial intelligence improves patient follow-up in a diabetic retinopathy screening program. Clin. Ophthalmol. 17, 3323–3330 (2023). sourcePartly supported

  The Discussion (page not provided) suggests that faster release of results and changes to patient messaging may improve follow-up. But the observed three-fold gain occurred with an AI workflow compared with manual-grading workflows; the study did not test workflow redesign alone while retaining manual grading. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

Two studies excluded ungradable images from analysis entirely, prior to allocation to AI or comparator arms.

“Two studies excluded ungradable images from analysis entirely, prior to allocation to AI or comparator arms14,18.”

• cites [14] Mathenge, W. et al. Impact of artificial intelligence assessment of diabetic retinopathy on referral service uptake in a low-resource setting: the RAIDERS randomized trial. Ophthalmol. Sci. 2, 100168 (2022). sourceSupported
• cites [18] Li, J. et al. Integrated image-based deep learning and language models for primary diabetes care. Nat. Med. 30, 2886–2896 (2024). sourcePartly supported

  The Methods section on DeepDR-Transformer as an assistive tool says only gradable eyes were selected for the assisted-versus-unassisted comparison, supporting exclusion from that comparison. But the Methods section on retrospective evaluation also reports an analysis of gradable versus ungradable images, so ungradable images were not excluded from analysis entirely. The comparison used the same cases sequentially before and after AI assistance, rather than allocating cases to separate arms; no page numbers are available in our copy. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

Three studies included ungradable images as referable DR, thus retaining them in the referral denominator.

“Three studies included ungradable images as referable DR, thus retaining them in the referral denominator15–17.”

• cites [15] Dow, E. R. et al. Artificial intelligence improves patient follow-up in a diabetic retinopathy screening program. Clin. Ophthalmol. 17, 3323–3330 (2023). sourcePartly supported

  The Results (page not available) count ungradable images among referable screening results in the Human and AI–Human Hybrid workflows: their referral denominators are 117 and 103, respectively. The AI Workflow denominator of 279 includes only MTMDR-positive results, so this paper does not show ungradable images retained in every referral denominator it reports. The paper also cannot establish the cited sentence’s count of three studies. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [16] Chotcomwongse, P. et al. Transforming non-digital, clinical workflows to detect and track vision-threatening diabetic retinopathy via a digital platform integrating artificial intelligence: implementation research. Ophthalmol. Ther. 14, 447–460 (2025). sourcePartly supported

  The Methods (Implementation and Variables; page unavailable) define an ungradable retinal photograph as a reason for referral, and the Results include ungradable cases among referred patients. The paper treats ungradable photographs as distinct from vision-threatening DR, however, so it does not support describing them as referable DR. The “three studies” count belongs to the citing sentence’s three-reference group, not to this source alone. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [17] Liu, J. et al. Diabetic retinopathy screening with automated retinal image analysis in a primary care setting improves adherence to ophthalmic care. Ophthalmol. Retin. 5, 71–77 (2021). sourcePartly supported

  The Methods say patients with inconclusive screening results were referred, and the Results say positive and inconclusive results were pooled when assessing abnormal screening results. But the Results list inconclusive results separately from referable diabetic eye disease; the paper does not classify those images as referable DR. The claim that three studies used this approach cannot be established from this source alone. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

The risk difference for referral uptake with AI-assisted versus non-AI referrals was 23.93% (95% CI 12.81% to 35.05%).

“and risk difference (RD) of 23.93%, 95% CI, 12.81%, 35.05% (Fig. 3) for referral uptake.”

Settings which underwent referral pathway transformation from routine to targeted referrals for diabetic retinopathy demonstrated the greatest effect size.

“Settings which underwent referral pathway transformation from routine to targeted referrals for DR demonstrated the greatest effect size.”

• AssumptionRests on an assumption: The included studies are sufficiently comparable in their definitions of referral uptake and artificial intelligence-assisted care pathways to be pooled in a meta-analysis.

There was substantial heterogeneity between studies (I2 = 91.9%).

“our analysis showed substantial heterogeneity between studies (high I² value, 91.9%)”

Most (n = 4) included studies utilised behavioural change interventions enabled by immediate results acquisition of artificial intelligence to enhance health-seeking behaviour.

“Most (n = 4) studies also utilised behavioural change interventions enabled by immediate results acquisition of AI to enhance health-seeking behaviour.”

The relative risk for referral uptake with AI-assisted versus non-AI referrals was 1.887 (95% CI 1.175 to 3.030).

“We calculated a relative risk (RR) of 1.887, 95% CI, 1.175, 3.030 (Fig. 2)”

Two studies compared AI-assisted targeted referrals to routine referral strategies where all diabetic patients were referred without results.

“Two studies compared AI-assisted targeted referrals to routine referral strategies where all diabetic patients were referred without results.”

The relative risk of diabetic retinopathy referral uptake with artificial intelligence-assisted screening compared with the status quo was 1.89 (95% CI 1.18, 3.03, I2 = 91.9%).

“The relative risk of DR referral uptake with AI-assisted screening compared with the status quo was 1.89 (95% CI, 1.18, 3.03, I2 = 91.9%).”

• AssumptionRests on an assumption: The included studies are sufficiently comparable in their definitions of referral uptake and artificial intelligence-assisted care pathways to be pooled in a meta-analysis.

Based on a random effects meta-analysis of six studies comparing AI-assisted versus non-AI referrals for diabetic retinopathy, AI screening improves referral uptake compared to traditional care pathways.

“Based on the random effects meta-analysis of six studies comparing AI-assisted versus non-AI referrals for DR, the pooled results suggest that AI screening improves referral uptake compared to traditional care pathways.”

There was inconsistency in the effect of timeframe and rates of referral uptake.

“There was inconsistency in the effect of timeframe and rates of referral uptake.”

A clear pattern emerged whereby studies with lower baseline rates of referral uptake demonstrated larger absolute improvements.

“A clear pattern emerged whereby studies with lower baseline rates of referral uptake demonstrated larger absolute improvements.”

Studies replacing routine referrals for all people with diabetes with AI-assisted targeted referrals in low-performing systems (baseline adherence 18.7-22.0%) achieved the largest improvements in referral uptake.

“Studies replacing routine referrals for all people with diabetes with AI-assisted targeted referrals in low-performing systems (baseline adherence 18.7–22.0%) achieved the largest improvements.”

There is substantial heterogeneity (I2 = 92%) between studies of AI-assisted DR screening.

“Despite the clinical significance, substantial heterogeneity (I² = 92%) indicates marked variability between studies and context-dependent effectiveness which will be discussed further.”

• Also stated in: § Discussion · plain

  “However, these findings are supported by low-certainty evidence due to serious inconsistency in effect sizes across studies, heterogeneity (I2 = 92%) and potential publication bias.”

• Also stated in: § Discussion · plain

  “We attribute substantial heterogeneity (I2 = 92%) to three principal sources.”

Four studies compared AI-assisted screening directly to manual grading workflows.

“Four studies compared AI-assisted screening directly to manual grading workflows.”

Comprehensive AI-assisted care pathways consistently improve DR referral uptake across diverse settings.

“comprehensive AI-assisted care pathways consistently improve DR referral uptake across diverse settings.”

Ungradable images were managed differently across studies, which warrants consideration as a potential source of bias.

“Ungradable images were managed differently across studies, which warrants consideration as a potential source of bias.”

AI-assisted screening pathways consistently improve referral uptake among patients with referable DR across diverse settings.

“AI-assisted screening pathways consistently improve referral uptake among patients with referable DR across diverse settings”

None of the observational studies adequately controlled for confounding variables.

“None of the observational studies adequately controlled for confounding variables.”

Evidence on clinical and health system outcomes such as referral uptake from real-world artificial intelligence-assisted diabetic retinopathy screening remains limited.

“Real-world implementation of DR screening with AI-assisted care pathways has accelerated globally since 20208,13–22, yet evidence on clinical and health system outcomes such as referral uptake remains limited.”

Overall risk of bias among the included studies was low to serious.

“Overall, the risk of bias among studies was low to serious.”

The effectiveness of AI versus non-AI referrals may depend on contextual factors such as healthcare setting, patient population, follow-up duration, and implementation approach.

“suggesting that the effectiveness of AI versus non-AI referrals may depend on specific contextual factors such as healthcare setting, patient population, follow-up duration, and implementation approach.”

AI-assisted workflows may have greatest impact in settings with currently poor rates of referral uptake, particularly where routine referral strategies are being replaced by targeted, screening-based approaches to diabetic retinopathy referral supported by AI.

“This suggests AI-assisted workflows may have greatest impact in settings with currently poor rates of referral uptake, particularly where routine referral strategies are being replaced by targeted, screening-based approaches to DR referral supported by AI.”

Attribution of improvements in completed referrals specifically to AI versus other workflow components remains challenging given the frequent inclusion of co-interventions and fundamental differences in referral strategies across studies.

“Attribution of improvements in completed referrals specifically to AI versus other workflow components remains challenging given the frequent inclusion of co-interventions and fundamental differences in referral strategies across studies.”

There are diminishing returns in settings with already functional screening infrastructure.

“This was most pronounced in in higher-performing systems (baseline referral uptake 39.6–77.3%)14,16,18 which indicates diminishing returns in settings with already functional screening infrastructure.”

The effectiveness of AI-assisted DR screening on referral uptake is context-dependent.

“Despite the clinical significance, substantial heterogeneity (I² = 92%) indicates marked variability between studies and context-dependent effectiveness which will be discussed further.”

The evidence for the effect of AI-assisted DR screening on referral uptake is of low certainty due to serious inconsistency in effect sizes across studies, heterogeneity (I2 = 92%) and potential publication bias.

“However, these findings are supported by low-certainty evidence due to serious inconsistency in effect sizes across studies, heterogeneity (I2 = 92%) and potential publication bias.”

For approximately every four patients screened through an AI-assisted pathway compared with the status quo, one additional patient completed their referral.

“In absolute terms, for approximately every four patients screened through an AI-assisted pathway compared with the status quo, one additional patient completed their referral.”

Settings likely to benefit most from AI-assisted screening pathways include those currently using routine referral strategies or lacking systematic screening infrastructure, where baseline rates of referral uptake are low (<30%).

“Settings likely to benefit most include those currently using routine referral strategies or lacking systematic screening infrastructure, where baseline rates of referral uptake are low (< 30%).”

This system-level transformation improved supply-side value.

“thereby improving supply-side value.”

The relative contribution of AI diagnostic capability, immediate results acquisition, and enabled co-interventions to referral uptake remains uncertain.

“The relative contribution of AI diagnostic capability, immediate results acquisition, and enabled co-interventions, therefore remains uncertain.”

We included both randomised and observational studies reflecting real-world implementation contexts.

“included both randomised and observational studies reflecting real-world implementation contexts”

AI-assisted screening improves referral uptake through patient-level behavioural interventions enabled by immediate results acquisition.

“AI-assisted screening appears to improve referral uptake through two mechanisms: patient-level behavioural interventions enabled by immediate results acquisition, and health system redesign by transforming referral pathways.”

Timeframe of referral may influence reported rates of uptake.

“Timeframe of referral may also influence reported rates of uptake.”

Such variability makes it difficult to ascertain the true effect of longer appointment timeframes on corresponding uptake.

“Such variability makes it difficult to ascertain the true effect of longer appointment timeframes on corresponding uptake.”

The improvement in referral uptake from AI-assisted DR screening is clinically significant.

“This is clinically significant given the progressive nature of DR and risk of blindness if left untreated through non-uptake of a referral.”

• AssumptionRests on an assumption: Diabetic retinopathy is progressive and can cause blindness if left untreated through non-uptake of a referral.

AI-assisted screening improves referral uptake through health system redesign by transforming referral pathways.

“AI-assisted screening appears to improve referral uptake through two mechanisms: patient-level behavioural interventions enabled by immediate results acquisition, and health system redesign by transforming referral pathways.”

AI's value proposition from a health system perspective depends on whether existing care pathways require transformation or optimisation.

“This pattern suggests AI’s value proposition from a health system perspective depends on whether existing care pathways require transformation or optimisation.”

Among included studies in our review, the risk from differential exclusion of ungradable images was mitigated by design.

“However, among included studies in our review this risk was mitigated by design (Table 2).”

Temporal confounding affects retrospective studies using historical controls, where concurrent quality improvement initiatives or evolving health system factors may independently influence referral uptake.

“Second, temporal confounding affects retrospective studies using historical controls, where concurrent quality improvement initiatives or evolving health system factors may independently influence referral uptake.”

Low-certainty evidence limits attribution of effects to specific pathway components.

“though low-certainty evidence limits attribution to specific pathway components.”

Pooling of randomised and non-randomised studies introduced complexity for interpretation.

“Pooling of randomised and non-randomised studies introduced complexity for interpretation”

Such heterogeneity may limit the reliability of our random effects meta-analysis.

“Such heterogeneity may limit the reliability of our random effects meta-analysis”

These limitations undermine causal inference about AI's isolated effect on referral uptake.

“These limitations undermine causal inference about AI’s isolated effect on referral uptake”

The included studies' diversity supports generalisability.

“supporting generalisability.”