Development of a Mobile Terminal-Based Intelligent Detection System for Multiple Anterior Segment Diseases of the Eye

Condition(s) studied

Investigational drug(s) / intervention(s)

Smartphone-based on-device artificial intelligence system for anterior segment eye disease screening: A structured-pruned one-stage object-detection model deployed as a standalone Android application, performing all image inference on-device without internet connectivity, designed to detect 16 anterior segment eye diseases from smartphone-captured images.

Study summary

This is a multi-center, cross-sectional study evaluating a smartphone-based artificial intelligence (AI) system for anterior segment eye disease screening. The system is designed to identify 16 clinically important anterior segment conditions from images captured using a standard Android smartphone. A core design feature of the system is that all image analysis is performed entirely on the smartphone itself, without requiring internet connectivity or cloud-based server infrastructure.

The study is motivated by a structural challenge in the deployment of medical AI: systems that depend on cloud infrastructure for inference are non-functional in settings without reliable internet access, which disproportionately excludes populations in low-resource regions where the burden of preventable eye disease is highest. This study evaluates whether an on-device AI system, designed with operational constraints as a primary engineering objective, can deliver clinically acceptable diagnostic performance while remaining operable under real-world connectivity limitations.

The study comprises five evaluation components. First, the diagnostic performance of the AI system is benchmarked against board-certified ophthalmologists of varying seniority on a standardized set of smartphone-captured anterior segment images. Second, the usability of the system is evaluated among non-medical users who perform self-administered screening with minimal instruction, with per-screening time recorded across consecutive attempts to characterize the learning curve. Third, a head-to-head field trial directly compares the on-device AI system against a functionally equivalent cloud-based deployment of the same model architecture across key operational dimensions including screening duration, diagnostic performance, and user acceptability. Fourth, population-level screening is conducted among consecutively enrolled community residents at two low-resource sites, with per-disease sensitivity and specificity calculated against reference-standard slit-lamp examinations. Fifth, pre-specified health-economic and environmental analyses compare the two deployment modalities in terms of per-person screening cost, cost-effectiveness, per-inference electricity consumption, and projected carbon emissions at scale.

The reference standard for all diagnostic comparisons is slit-lamp biomicroscopic examination performed by board-certified ophthalmologists. The study is designed and reported in accordance with the DECIDE-AI reporting guideline for early-stage clinical evaluation of AI-driven decision-support systems.

Eligibility

Primary outcome measure(s)

• Case-level diagnostic accuracy of the AI system compared with board-certified ophthalmologists — Day 1
  Case-level accuracy is defined as the proportion of images with fully correct diagnostic labels concordant with the reference standard. The AI system and board-certified ophthalmologists stratified by clinical seniority (junior: fewer than 5 years of independent practice; intermediate: 5 to 15 years; senior: more than 15 years) independently evaluate the same standardized set of smartphone-captured anterior segment images, sampled to ensure balanced representation of all 16 disease categories and normal eyes. The reference standard for each image is established by a senior ophthalmologist with more than 30 years of clinical experience who does not participate in the benchmarking exercise. Clinicians are masked to the AI system output and to each other's assessments throughout.
• Diagnostic accuracy of the AI system when operated by non-medical users — Day 1
  Diagnostic accuracy is defined as the proportion of images correctly classified by non-medical users operating the AI system independently. Non-medical users, including patients and their family members attending the outpatient clinic of Zhongshan Ophthalmic Center, are instructed to install the AI application on their own smartphones, follow the in-app guidelines, and capture anterior segment images of an accompanying person to receive a screening result. All inference is performed on-device without internet connectivity. AI-generated diagnostic outputs are compared with reference-standard diagnoses obtained from subsequent slit-lamp examinations performed by ophthalmologists at the same clinic.
• Sensitivity of the on-device AI system in population-level community screening — Day 1
  Sensitivity is defined as the proportion of participants with a given anterior segment disease who are correctly identified as positive by the on-device AI system (true positives divided by the sum of true positives and false negatives), calculated separately for each target disease category. Village staff without medical training use the on-device AI system to screen consecutively enrolled local residents. The reference standard is established through slit-lamp examinations performed by ophthalmologists.
• Incremental cost-effectiveness ratio of on-device versus cloud-based screening — Day 1
  The incremental cost-effectiveness ratio (ICER) is defined as the difference in lifetime costs between on-device and cloud-based screening divided by the difference in quality-adjusted life-years (QALYs) between the two strategies, estimated from a pre-specified decision-analytic model comprising a decision tree with a downstream Markov state-transition structure applied to a simulated cohort of 100,000 individuals.
• Per-inference electricity consumption of on-device versus cloud-based deployment — Day 1
  Electricity consumption in joules per inference cycle is measured over 10,000 inference cycles on both the on-device smartphone and a standardized cloud-server configuration (Intel Xeon Gold 6248 CPU, NVIDIA Tesla V100 GPU) using the Experiment Impact Tracker toolkit. Cloud-server measurements are averaged across low-traffic and high-traffic server conditions. Results are reported separately for the on-device and cloud-based deployment modalities and compared as a ratio.

Trial sites (1)

More Zhongshan Ophthalmic Center, Sun Yat-sen University trials in China

Other trials for the same condition