AI-automated referral for patients with visual impairment

In the past 5 years, deep neural networks have been applied in several medical applications such as detecting pneumonia and lung consolidation from chest x-rays CT scans,1Zhang K Liu X Shen J et al.Clinically applicable AI system for accurate diagnosis, quantitative measurements, prognosis of COVID-19 using computed tomography.Cell. 2020; 1821360Summary Full Text PDF PubMed Scopus (19) Google Scholar melanoma skin photographs.2Esteva A Kuprel B Novoa RA al.Dermatologist-level classification cancer with networks.Nature. 2017; 542: 115-118Crossref (54) Deep learning has also integrated into ophthalmic applications, especially use fundus optical coherence tomography images because their widespread clinical settings. shown to detect diabetic retinopathy, glaucoma, age-related macular degeneration, oedema, retinopathy prematurity, even predict systemic cardiovascular risk factors.3Rim TH Lee G Kim Y al.Prediction biomarkers retinal photographs: development validation deep-learning algorithms.Lancet Digit Health. 2: e526-e536Summary (24) Additionally, models be effective referrals tertiary care centres.4Dunnmon JA Ratner AJ Saab al.Cross-modal data programming enables rapid machine learning.Patterns (N Y). 8100019Summary (9) Scholar, 5Bellemo V Lim ZW al.Artificial intelligence screen referable vision-threatening Africa: a study.Lancet 2019; 1: e35-e44Summary (102) 6Wu Huang Z al.Universal artificial platform collaborative management cataracts.Br Ophthalmol. 103: 1553-1560Crossref (41) 7Brown JM Campbell JP Beers al.Automated diagnosis plus disease prematurity convolutional networks.JAMA 2018; 136: 803-810Crossref (244) Studies multiple pipelines triage x-rays;4Dunnmon referral (vision threatening),5Bellemo cataract,6Wu prematurity;7Brown telemedicine programmes photographs. April, 2018, US Food Drug Administration approved first AI-based device retinopathy.8Abràmoff MD Lavin PT Birch M Shah N Folk JC Pivotal trial an autonomous diagnostic detection primary offices.NPJ Med. 39Crossref The Lancet Digital Health, Yih-Chung Tham colleagues9Tham Y-C Anees Zhang L al.Referral disease-related visual impairment photograph-based learning: proof-of-concept, model 3: e29-e40Google report single modality (fundus photograph) algorithm goal creating vision care. Following recommendation WHO, some high-income countries implemented annual screening, older populations.9Tham Some screening systems rely on acuity checks whereas others photography identify To incorporate referral, large-scale resources trained personnel are required. However, robustly could potentially automate this process. develop algorithm, colleagues used Singapore Epidemiology Eye Disease Study (SEED) train model, they validated five external datasets. diverse datasets contained eyes Malay, Indian, Chinese White participants, mean age across ranging 48·2 years (SD 13·4) 76·6 (6·8). is two-tiered, two separate algorithms prediction presence (classification task), best-correlated level (regression task). authors standard network (Residual Neural Network [ResNet]-50) regression. predictor new. ResNet did well predicting any impairment, area under receiver operating characteristic curve (AUC) range varying 86·6% (95% CI 83·4–89·7) 93·6% (92·4–94·8) datasets.9Tham datasets, AUC mild was 81·9% (77·2–86·6) 92·5% (90·8–94·2) moderate 85·9% (81·8–90·1) 93·5% (91·7–95·3). According these data, quite indicating potential generalisability, which crucial element deployment utility. visualise regions interest generated saliency maps GradCAM method. They found that were congruent pathological signs correlated impairment. Interestingly, when more than one present image, had ability all features different contributory weights. One limitations developed by paucity representation populations. Depending where deployed, it needs further representative population. Another limitation high class imbalance (approximately 10% positive labels). mentioned weight training but better method might encompass sample general Furthermore, similar other systems, susceptible bias towards ground labels learned only what presented images, not represent real-world data. authors' approach eye-care enable eye substantial Subsequent studies include specific stages causing loss. These valuable standpoint, serving tool both referral. We observed robust identifying diseases stages. For example, how can control versus whether mild, moderate, or severe non-proliferative proliferative retinopathy.10Gulshan Peng Coram al.Development photographs.JAMA. 2016; 316: 2402-2410Crossref (3079) flag either hopeful future multidisease able early stage then Despite exponential growth medicine ophthalmology, translations practice remain challenging. Algorithms need generalisable populations explainable ensure targeted performance, population representation, identification problems implement patients collaboration between researchers health-care workers centres, regulatory agencies, forged successful integration within workflow. JAH received grant funding support BrightFocus Foundation outside submitted work. declare no competing interests. Referral studyThis proof-of-concept study shows single-modality, function-focused related major diseases, providing timely pinpointed community hospitals. Full-Text Open Access