Is AI ready for low-vision rehabilitation?
A registered review found impressive navigation, recognition and accessibility prototypes across 19 studies. Yet only four studies measured a user-facing functional or usability outcome, and none used a validated quality-of-life or functional-vision instrument.
The short version
3 key points
- 1.The prospectively registered review searched four databases to 27 February 2026, screened 2,420 unique records and included 19 peer-reviewed studies.
- 2.Four studies reported a user-facing functional or usability outcome; the other 14 reported technical performance without a corresponding patient-reported or functional outcome.
- 3.Promising accuracy figures came from different tasks and settings and cannot be combined into one estimate of clinical effectiveness.
The answer: technically promising, clinically unproven
AI systems can already identify objects, read signs, map routes and suggest low-vision aids under selected test conditions. A new systematic review found examples of sub-metre localization, high object-recognition accuracy and successful navigation tasks. Those figures show that engineers can build useful components. They do not establish that a person with low vision will be safer, more independent or better able to work and study after adopting a system. The review’s most important finding is the distance between benchmark performance and patient-important evidence.
Across the 19 included studies, only four incorporated a user-facing functional or usability outcome. Fourteen reported technical metrics without a matching functional or patient-reported outcome, and no study used a validated quality-of-life or functional-vision instrument. That makes the review a map of prototypes and early evaluations, not a clinical verdict. Clinicians and commissioners should resist translating an accuracy percentage into a promise of independence unless the underlying task, test environment, error costs and participant experience are clear.[1]
How the review was conducted
The researchers registered the protocol prospectively in PROSPERO as CRD420261324781 and followed PRISMA reporting guidance. They searched PubMed, Scopus, Web of Science and IEEE Xplore from inception to 27 February 2026. After removing duplicates, 2,420 records were screened; 34 full texts were assessed and 19 peer-reviewed empirical studies were included. Two reviewers independently screened records and extracted information. Eligible designs ranged from randomized and quasi-experimental work to cohort, usability, prototype and mixed-methods studies.
That breadth was appropriate for an emerging field but prevented a meta-analysis. Navigation, object detection, reading, device fitting and interface repair measure different things on incompatible scales. The authors therefore synthesized results narratively across six application domains. They did not complete a formal study-quality appraisal because the included papers mainly described system development and used dissimilar technical measures. The review itself was funded by the Else Kröner-Fresenius-Stiftung Foundation and the NKABOM project; the authors declared no commercial conflict.[1]
What worked—and under which conditions
Navigation was the largest domain, represented by eight studies. One UNav evaluation reported 100% task success and fewer wrong turns than unassisted navigation, but its server-dependent pipeline added about 3.62 seconds of round-trip delay. An augmented-reality sign-reading system improved route efficiency while users took roughly 129 seconds longer to complete tasks. Other systems performed strongly on sign detection, curbstone segmentation or indoor localization, yet their test conditions varied and some results came from held-out datasets rather than live journeys.
Recognition systems produced similarly mixed signals. Reported object or scene accuracy could be high for well-defined, high-contrast categories while falling for complex or low-contrast inputs. Construction-hazard detection reached 92% dynamic accuracy after temporal filtering, but motion blur, oblique viewpoints and camera limitations still caused failures. A chart-description system for educational videos achieved an F1 score of 86.1% for detection, yet generated captions could hallucinate and did not handle some complex charts. Those are meaningful engineering results with safety-critical exceptions.[1]
The real-world gap is not a footnote
Lighting degraded performance in five studies. Rain, fog, snow, clutter and changing illumination also challenged some systems. Cloud processing introduced dependence on connectivity; on-device processing reduced that dependence but brought battery, heat and computational constraints. Text-reading pipelines took 31 to 66 seconds in some configurations, while optical-character recognition could add three to four seconds per sign. A delayed or missing description may be inconvenient in a document and dangerous at a road crossing, so average accuracy is not enough.
Representation is another uncertainty. Dataset sizes were inconsistently reported, and few studies clearly documented evaluation on a truly independent dataset. Some systems were tested by sighted people in simulated low-vision conditions or evaluated through expert ratings rather than everyday use by the intended population. The review excluded company deployments and unpublished work, so it may miss operational products; conversely, academic publication does not guarantee that a prototype is available, affordable, maintainable or compatible with a person’s preferred rehabilitation plan.[1]
What clinicians, developers and users can do now
Low-vision services can treat AI tools as candidates for individualized trials rather than replacements for established rehabilitation. A useful assessment should specify the person’s goals, residual vision, mobility context, device access and tolerance for audio, visual or haptic feedback. Staff should test the tool in the environments where it will be used, document failure modes and retain non-AI alternatives. Procurement questions should cover offline operation, update policy, privacy, battery life, incident reporting and what happens when the model is uncertain.
Developers should involve people with low vision from problem definition through evaluation. Safety-critical functions need on-device fallbacks, clear confidence cues and an easy way to stop or override guidance. Performance should be disaggregated by lighting, weather, device, condition and user group. Affordability and training matter as much as model architecture: a system that assumes a recent phone, continuous data connection or precise camera alignment can widen access gaps even when its laboratory score is excellent.[1]
What would change the assessment
Confidence would rise with prospective comparative studies that recruit diverse people with low vision and measure real tasks over months, not minutes. Common outcomes should include falls or near misses, route completion, reading performance, fatigue, device abandonment, confidence, participation and validated functional-vision and quality-of-life scores. Trials should report adverse events, false reassurance and the burden of checking incorrect outputs. Independent replication outside the developers’ own dataset and site is essential.
Until that evidence arrives, the most defensible conclusion is narrower than the strongest performance figures. AI can make parts of low-vision support more adaptive and context-aware, and several prototypes deserve careful field testing. The review does not show that AI rehabilitation improves health, independence or quality of life at population scale. People choosing a tool should receive both the technical promise and the practical uncertainty, with clinicians and users—not benchmark leaderboards—deciding whether the trade-off is acceptable.[1]
What this means for people
- People with low vision should be shown how a tool fails in the environments where they will use it, not only its best accuracy figure.
- Clinicians can trial AI alongside established rehabilitation while retaining non-AI fallbacks and documenting user-important outcomes.
- Commissioners should require accessibility, offline resilience, privacy and post-deployment incident evidence before large purchases.
Global context
The review team spans Ghana, France, Germany and the United States, but the included systems and test settings were highly varied. Affordability, connectivity and rehabilitation provision differ sharply across countries, so a prototype’s portability cannot be assumed.
What the evidence does not yet show
- Only 19 heterogeneous studies and no quantitative meta-analysis
- Four user-facing functional or usability evaluations; 14 technical-only evaluations
- No validated quality-of-life or functional-vision outcome in the included studies
- Inconsistent dataset reporting and limited independent validation
- Published academic studies may omit commercial deployments and negative unpublished results
What to watch next
- Prospective comparative trials with people who have low vision
- Validated functional-vision, participation and quality-of-life outcomes
- Safety testing across lighting, weather, clutter and connectivity failures
- Independent datasets and external clinical sites
- Affordability, abandonment, privacy and post-deployment incident reporting
Living evidence record
Impact record IAI-0FFJNUN
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
15 October 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what Frontiers in Digital Health published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 15 October 2026
Research question, topics and regions
Whether AI-enabled navigation, recognition and accessibility systems have evidence of real-world benefit for people with low vision
Topics: Low vision · Assistive technology · Computer vision · Rehabilitation · Accessibility
Regions: Ghana · Africa · Europe · United States
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Watch the background, then check the evidence
The Royal Society
AI-generated illustration · not a video frame
YouTube loads only when you choose. Press play in the player to start. If playback is unavailable, use “Watch on YouTube” below.
Healthcare, culture & human impact · Background explainer
Published 2026-07-01 · 44 min 28 sec
Is this the AI future we want? | The Royal Society
A Royal Society discussion with Madhumita Murgia and biomedical engineer Alison Noble on AI’s potential in healthcare, science and culture, recorded at the 30 June 2026 event.
Evidence check: A discussion of opportunities and risks, including future possibilities. It is not evidence that every application described is available or clinically proven.
Source checked 14 October 2026. Check available captions on the publisher’s video page.
The Impact Brief · Free
Keep following the stories that matter to you.
Get our weekday briefing on what changed and why it matters. Choose your topics below.
Continue the story
There’s more to this story
Health & Life Sciences
Can endoscopy AI travel beyond still images?
A new classifier reached about 94% accuracy in cross-validation and 96% on an aligned external image subset, then narrowly outscored five gastroenterologists on 200 still images. Both datasets came from Norway, and the work did not test live video, workflow or patient outcomes.
8 min · 1 source
Health & Life Sciences
Can AI identify a medicine safely outside the lab?
A review of 91 studies finds rapid progress in pill detection, classification and imprint reading, but little external, prospective or clinical validation. High benchmark accuracy is not evidence that an app prevents medication harm.
7 min · 1 source
Health & Life Sciences
Does 93.5% Dice make this ovarian-cancer CT AI clinically ready?
No. A two-stage ensemble improved segmentation on a patient-separated TCGA-OV test set, but the study involved only 143 patients, inherited molecular labels at patient level and had no external or prospective clinical validation.
8 min · 2 sources
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.