A recent review, published in Current Opinion In Ophthalmology, explores the current status of artificial intelligence (AI) systems in ophthalmology. It also illustrates the steps required for clinical translation of AI into personalised healthcare for retinal diseases.
The potential of AI in ophthalmology
In recent years, the potential applications of AI, particularly deep learning, to transform healthcare have become increasingly apparent. At the forefront of this transition has been ophthalmology. Experts hope that utilising predictive algorithms and harnessing the power of data will be a big step towards personalised healthcare. Most importantly, they expect that AI-assisted scientific discovery may provide novel insights into disease pathophysiology and help improve clinical trial efficiency. While there have been rapid advances in AI-enabled healthcare in retinal disease, there are still several challenges that remain before real patient benefit can be seen. The authors note that despite the hype, AI in ophthalmology is still in its early stages. They believe more work needs to be done before AI deployment is scalable, provides meaningful real-world application and achieves human interaction.
For training and evaluation of deep-learning models, high-quality clinical datasets are essential. The authors suggest that harnessing advanced analytic methods on large real-world datasets is important for this. Training sets must be representative of all populations. Additionally, they must be in a computationally tractable form, appropriately annotated and robustly deidentified. However, meeting these requirements can be challenging and time consuming. In ophthalmology, there has been an increased use of wider datasets derived from routine clinical care. For example, the American Academy of Ophthalmology IRIS Registry of electronic health records (EHRs). Investigators should make large datasets available to allow independent benchmarking of AI systems.
The team highlight that advances in technology will result in the expansion of data sources and modalities. For example, increased usage of 5G technology will allow individuals to capture personal sensor and hyperlocal environmental data. Investigators could then use this data as the substrate for AI models. Additionally, there will be an accumulation of ‘omic’ data. Therefore, the development of systems and infrastructure to integrate and manage this data will also be important.
AI systems in ophthalmology
Convolutional neural networks have proven effective for image classification tasks. As a result, most early work of AI in ophthalmology focused on classification of retinal photographs/COT images. As the AI field continues to evolve and datasets grow, it is likely that the range of applications for AI will greatly expand. The authors emphasise that as the translational pipeline for clinical AI systems evolves, identification and interrogation of potential use cases will be crucial.
Robust validation of an algorithm’s performance is critical. AI systems must have established benchmarks for human diagnostic performance at a range of expertise levels for screening, diagnosis and future prediction of retinal disease. Most importantly, when used, AI systems must result in improvements or be at the same level as human diagnostic performance.
While researchers have used several retrospective studies to train and test deep-learning algorithms, very few prospective studies have evaluated AI performance in ophthalmology. Examples of validation studies could include prospective observational validation by exposure of the algorithm to data representative of real-life diversity and also prospective randomised controlled trials (RCTs). Prospective studies are necessary to understand the diagnostic accuracy of AI systems in real-world settings. Specifically, intervention RCTs are needed to address issues of clinical effectiveness.
Technical maturity of AI systems will depend upon performance validation in the general population. As a result, the authors suggest that frameworks and metrics for AI performance reporting will need to be established. They also emphasise that regulatory frameworks are fundamental to achieve safe and effective deployment of AI algorithms. Regulatory requirements may vary based on the particular use case. Nonetheless, it is critical that key stakeholders involved in the development and implementation of AI systems work together with regulators in developing the regulatory framework. Methods used to monitor safety must detect both underlying technology and wide system issues. Adoption of AI systems is currently largely hindered by the perception that AI is a ‘black box’. Therefore, improving interpretability of AI systems to interrogate the decision-making process will also be key.
In addition, implementation and sustainability of AI systems will require a financially viable business model. Therefore, the authors suggest that researchers should conduct future research, evaluating cost-effectiveness of AI systems in large cohorts across multiple AI applications in healthcare and within specific use cases. Another issue of AI is the lack of laws surrounding medical liability involving AI. As AI moves into clinical practice, standards of care around AI must be established.
For personalised healthcare, collaboration of all stakeholders will be key in tackling existing challenges. While AI is showing potential in areas like ophthalmology, further work needs to be done. This includes aggregating datasets, training and validating AI systems, establishing regulatory frameworks, AI implementation and adoption model adjustment and meaningful human-AI interaction. The authors believe that AI will allow ophthalmologists to gain insight from large volumes of multivariate data and interpret AI recommendations within a clinical context. In turn, the field of ophthalmology will be in a position to lead transformation of healthcare towards personalised care.
Image credit: By starline – www.freepik.com