Artificial Intelligence Applications and Cataract Management: A Systematic Review

Authors: Tognetto D, Giglio R, Vinciguerra AL, Milan S, Rejdak R, Rejdak M, Zaluska-Ogryzek K, Zweifel S, Toro MD.

Geographical coverage: Not reported

Sector: Service Delivery

Sub-sector: Screening

Equity focus: Not reported

Study population: Patients with cataract

Review type: Effectiveness review

Quantitative synthesis method: Meta-analysis

Qualitative synthesis method: Not applicable

Background: Cataract is a major global cause of visual impairment, accounting for more than 50% of blindness cases in low- and middle-income countries. Cataract surgery is highly cost-effective, yielding improvements in patients’ physical and psychological well-being and even enhancing cognitive function in those with dementia. However, significant challenges persist, especially in public health systems, including limited resources and long waiting times for surgery. The COVID-19 pandemic further disrupted timely diagnosis and treatment of cataracts. Amid these challenges, artificial intelligence (AI) has rapidly entered the healthcare field and shown promising results in screening, diagnosis, and treatment. With ageing populations and expanding digital infrastructure, AI has the potential to significantly reshape cataract management.

Objective

To evaluate the various applications of AI-based software across all aspects of cataract management, including diagnosis, surgical planning, and follow-up care.

Main findings

The review included 49 studies and categorised the findings into four main areas of cataract management influenced by AI: diagnosis, intraocular lens (IOL) power calculation, surgery, and complications.

AI demonstrated significant potential in cataract detection and grading, often matching or surpassing the accuracy of experienced clinicians. Various studies utilised machine learning and deep learning techniques to analyse slit-lamp and fundus images for cataract identification. For instance, Fan et al. (2003) achieved 95.8% accuracy in cataract grading using slit-lamp photographs, while Wu et al. (2019) developed a universal AI platform that achieved an area under the curve (AUC) of over 99% for diagnosing cataracts. Other studies explored the use of non-mydriatic (no pupil dilation) fundus imaging and mobile smartphone applications, enabling remote monitoring and community-based screening programs. Nonetheless, challenges were noted, such as image noise (especially in fundus photographs) and the critical need for high-quality, diverse training data to improve AI model robustness.

AI has improved the accuracy of IOL power calculations, which are critical for achieving desired postoperative refractive outcomes in cataract surgery. Traditional calculation formulas have been enhanced by AI-driven methods like the Hill-RBF and Kane formulas, which generally show lower prediction errors compared to conventional formulas. For example, Ladas et al. (2021) reported reduced mean absolute error in refractive predictions using an AI-enhanced formula. However, the advantage of AI is not uniform in all cases; variability in results was observed across eyes with different axial lengths, and in some instances AI provided only marginal benefits over refined regression models. These findings suggest that while AI tools are promising, further refinement and validation are needed to consistently outperform traditional methods in all patient groups.

AI applications in cataract surgery have encompassed operating room logistics, real-time surgical guidance, and robotic assistance. One study by Devi et al. (2012) used AI algorithms to predict the duration of cataract surgeries, which helped in optimising operating room scheduling and efficiency. Other work, such as that by Quellec et al. (2014), involved video analysis systems that automatically segmented and recognised phases of cataract surgery. These systems can be used for surgical training and feedback, by identifying steps where a trainee might need improvement. Additionally, robotic and computer-assisted systems have been explored to enhance surgical precision and safety. The overarching aim of these technologies is to support surgeons (particularly those in training) with real-time decision support and to reduce the likelihood of human error or complications during surgery.

AI has also been employed to predict and manage postoperative complications of cataract surgery, such as posterior capsule opacification (a common long-term complication where the lens capsule becomes cloudy) and macular oedema. For instance, Mohammadi et al. (2012) used neural network models to predict which patients might require a future YAG laser capsulotomy to treat posterior capsule opacification. In another example, Hecht et al. (2019) developed an algorithm to help differentiate between macular oedema due to diabetes and that due to cataract surgery (pseudophakic macular oedema), which is important for appropriate management. Furthermore, natural language processing (NLP) techniques have been applied to analyse surgical reports and electronic health records, automatically extracting information about complications. This can streamline post-surgery monitoring and facilitate large-scale analysis of outcomes to improve care quality.

Methodology

The review’s literature search was conducted using PubMed, focusing on research articles that investigated AI applications in adult cataract management. Studies published from the inception of PubMed up to 1 March 2021 were considered. Only original studies involving human adult subjects were included. The reference lists of all included articles, as well as relevant previously published reviews, were scanned to identify any additional publications.

The study selection and data extraction were performed by three reviewers working independently. They screened titles and abstracts against the inclusion criteria, reviewed full texts, and extracted data on study characteristics and findings. In cases where there were disagreements or uncertainties, a fourth reviewer was consulted to resolve the issue. The findings from the included studies were synthesised narratively by the authors of the review, grouped into thematic areas (diagnosis, IOL calculation, surgery, complications) as described above, rather than being combined quantitatively.

Applicability/External Validity

While the results of this review highlight numerous promising AI applications in cataract care, the authors noted that the real-world applicability and generalisability of many AI algorithms remain limited. This is due to various challenges in their development and validation. For example, AI models often perform well in controlled research settings but may face difficulties when deployed in different hospitals or patient populations if the data used to train the models is not sufficiently diverse. The review emphasised that before these AI tools can be widely adopted in clinical practice, they need to be evaluated in robust clinical trials (preferably randomised controlled trials) to assess their safety, cost-effectiveness, and efficacy in real-world healthcare settings. In other words, it is crucial to determine whether using AI truly improves patient outcomes or workflow efficiency in a way that justifies any additional costs or training requirements. Only with such evidence can guidelines be developed for integrating AI into routine cataract management.

Geographical Focus

The review did not impose any geographical limitations on which studies were included, implying that it considered research from around the world. However, it did not explicitly report the geographical distribution of the studies it included. Thus, it’s unclear if the evidence base is concentrated in certain regions (such as East Asia, Europe, or North America where a lot of ophthalmology and AI research occurs) or truly global. The lack of reporting on study locations means one should be cautious in assuming the findings are universally applicable, as healthcare settings and patient demographics can differ widely across regions.

Summary of quality assessment: Overall, there is low confidence in the review’s conclusions. Inclusion and exclusion criteria were clearly defined. The screening and data extraction were performed independently by three reviewers, and reference lists of identified studies were scanned. Study characteristics were comprehensively presented, and findings were synthesised narratively to provide a cohesive overview. However, the literature search was restricted to PubMed, and the review did not assess study quality or risk of bias. Additionally, lists of included or excluded studies were not provided, and any potential language restrictions were not specified.