In recent years, Artificial Intelligence (AI) has grown to be an integral part of our society, utilised in our business, education and transport sectors. The integration of AI into the healthcare system has been a major talking point among many healthcare professionals, especially its role in healthcare, how this is expected to evolve in the future and also the associated benefits and consequences.
The efficiency, accuracy and effectiveness of AI and machine learning to diagnose disease, especially in its early stages compared to human efforts has been extensively studied. Applications for AI and machine learning within Ophthalmology has become a primary area of research and a number of very interesting studies have been published.
The summaries of four particularly insightful and relevant articles are available below, with a link to the full paper provided.
Artificial Intelligence (AI): Quo Vadis?: (Marco A. Zarbin.)
Why will AI influence ophthalmology and optometry?
AI may be able to do as well, or better than humans when it comes to image analysis. This is because computer vision uses information in each of the millions of pixels that may comprise an image. This is especially important as image analysis plays a central role in the diagnosis and management of conditions including glaucoma, Diabetes-related Retinopathy (DR) and Age-related Macular Degeneration (AMD).
AI may also have an important direct impact on access to care, monitoring, and treatment of chronic conditions and clinical trials.
AI has been shown to identify common vision threatening diseases with specificity and sensitivity that is comparable to that of experienced clinicians and AI may even predict DR progression.
This may allow for routine screening to be done outside the doctors’ offices. Therefore, clinicians can spend more time treating rather than screening patients. AI may also enable clinicals trials to be shorter with fewer patients as greater analytical capacity may improve hypothesis generation and patient selection, as well as support more sensitive clinical outcome monitoring.
Due to the reasons mentioned above, the use of AI may indirectly increase the amount of innovation in ophthalmology, optometry and vision science by decreasing the amount of time spent screening patients and increasing the amount of time clinicians can dedicate to research, improving data analysis.
What are the expectations and limitations of AI?
AI can revolutionize our interactions with electronic health records and improve the analysis of complex datasets.
Using AI voice recognition, natural language processing, and machine learning paradigms, could introduce the patient to clinicians as you enter the examination room and could list the relevant diagnoses, relevant changes in medical status, most recent treatment received by the patient, and current clinical impression based on an automated analysis of all the imaging data collected thus far during the visit.
AI might enable us to complete our professional tasks more efficiently, but also at higher level of competence.
However, it is highly unlikely that AI will replace physicians or scientists soon.
AI may be able to make treatment recommendations but it does not have the ability to answer a wide variety of questions a patient might have regarding treatment nor can it explain the potential complications of treatment in a manner that is clear and also appropriate for a wide variety of situations.
In conclusion, AI is becoming an integral part of our lives. Although AI will not replace scientists and clinicians in the near future, it can serve as a highly competent partner in our missions, a partner that will enable us to perform better, possibly at the highest level of which we are capable.
Will Artificial Intelligence (AI) Replace Ophthalmologists? (Edward Korot et al.)
Teamwork, creativity, adaptability and empathy are traits that physicians employ daily to effectively deliver patient care. However, although AI may mimic certain elements of human behaviours, there has not yet been a demonstration of empathy by an AI algorithm. This innately human doctor-patient relationship which AI lacks has been shown to have a therapeutic effect in patients.
Also, the translation of AI from laboratory experiment to real-world tool entails additional challenges. For example, although early studies showed that computers working with radiologists led to better accuracy than radiologists alone, subsequent clinical trials demonstrated that false-positive rates increased after computer aided diagnosis adoption. This led to an almost 20% increase in the rate of biopsies, confirming the potential disconnect between diagnostic accuracy and clinical effectiveness.
In ophthalmology, an ophthalmologist may not know whether the patient predicted to develop proliferative Diabetes-related Retinopathy, who was subsequently treated with a preventative anti-Vascular Endothelial Growth Factor (VEGF) injection, will have ever developed the disease if he or she was not treated. In this case false positive rates would be difficult to detect.
Novel uses of AI
AI have not only shown levels of performances that may supersede human ophthalmologists but have also demonstrated proficiency in tasks that were not previously thought possible for ophthalmologist to perform. The most striking demonstration to date can be seen in a deep learning algorithm that accurately predict cardiovascular risk factors and demographics from fundus photos. Further advances in AI-based ophthalmic image analysis will demonstrate unforeseen disease associations and their ophthalmic correlates which will not only enable earlier systemic disease detection but also novel insights into the pathophysiology of ophthalmic and systemic diseases.
Limitations of deep learning in ophthalmology
Although there is success in the use of deep learning-based AI models in research settings, there are shortcomings in real-world use, including insufficient transparency, poor integration with prior hierarchical knowledge, and inflexibility. AI may also break down when they encounter dissimilar image acquisition and patient-specific variables from those the model was trained on.
The potential of an Ophthalmologist-AI Partnership
Humans working with AI especially when looking at large data sets will be beneficial. However, these models can only lead to effective clinical decisions if they keep human intelligence “in the loop” to bring context.
Ophthalmology is both uniquely positioned to take advantage of AI such as integrating the ever-increasing volumes of clinical, genomic and imaging data, yet also uniquely protected against obsolescence to machines.
Humans (and human ophthalmologists) are underrated
Patients often complain of insufficient interaction with their doctor, physicians are also burnt out from more time spent on clerical tasks than patients. If implemented properly AI is unique in its potential to save time by processing large longitudinal data volumes and efficiently representing the patterns identified. This will allow ophthalmologists to have more time for physical contact and to focus on providing effective and compassionate care.
How artificial intelligence (AI) can transform randomised controlled trials – Cecilia S. Lee and Aaron Y. Lee
Randomised controlled trials (RCT) have been traditionally accepted as the most robust method of assessing the risks and benefit of any intervention. However, RCT is not always feasible due to rarity of disease, or time and costs that would impinge on the healthcare system.
The application of AI in RCTs may become a reality soon. Common shortcomings of unsuccessful RCTs include poor patient selection, inadequate randomization, insufficient sample size, and poor selection of end points.
With large datasets that incorporate clinical and multimodal imaging, AI models can be trained to select the potential study participants without relying on costly manual review to predict the natural history of each study participant, thus lowering the burden of individual screening and need for large sample sizes while selecting the patients who meet precise selection criteria.
AI will also help with avoiding potential confounders or misclassifications and allow shorter duration of RCTs due to the small sample size required.
The use of AI to select fast progressors alone will limit the generalizability of trial results and may expediate the development of novel therapies, for rare diseases.
Using AI in RCTs have the potential to minimise measurement errors and analyse the data without human-imposed biases.
Furthermore, AI models could generate new functional endpoint using structural data (e.g. OCT angiography from OCT) unlocking the potential of already archived data.
A challenge in RCT is a sufficient enrolment of patients who meet the inclusion and exclusion criteria. With enough data, AI models can be trained to predict the natural history of each participant. AI can allow for virtual controls; this cuts down recruitment significantly.
AI can be used to predict disease progression. This may increase the participation rate of subjects who are reluctant about participation due to the possibility of being in the placebo arm.
However, many limitations still exist with this class of machine learning algorithms. The quality of algorithms is heavily dependant on the availability of large, well-labelled data, which may not be free from measurement error.Methods that explore the source of an AI decision tree will be key in integrating AI into RCTs.
In conclusion, AI has the potential to improve and complement RCTs significantly in the future. However, AI will not replace RCTs. The synergy among clinicians, researchers, and industries in collaborative efforts to share and collect standardised data and allow AI algorithms to play major roles in RCTs may require paradigm shifts.
These efforts will expedite the development of AI in ophthalmology, which will ultimately increase the quality of care that is provided for individual patients.
Applications of Artificial Intelligence to Electronic Health Record Data in Ophthalmology (Wei Chun Lin).
The inclusion of electronic health records (EHR’s) into clinical practice has led to the collection of large quantities of clinical patient data, which is expected to identify ways to improve the quality of patient care, accuracy of disease diagnosis and risk assessment, particularly in ophthalmology.
However, significant challenges exist regarding the secondary use of EHR data and accurate data interpretation, especially when considering the diversity and quality of this data e.g. demographic information, eye exams and surgical records (partially complete, incorrect data entry/recording errors).
In recent years, Artificial Intelligence (AI) techniques have been used extensively to interpret data and provide new and more precise ways of diagnosing disease, predicting disease progression and risk assessment. In particular, AI techniques have been applied to examine their accuracy in detecting and predicting glaucoma, diabetes-related retinopathy, age-related macular degeneration (AMD) and cataracts, compared to traditionally-used methods.
Studies identified three common AI techniques used; machine learning, deep learning and natural language processing (NLP).
Machine learning exists in two forms; supervised and unsupervised machine learning. Supervised machine learning uses a model that has input and output data and predicts the output for new cases was most prominent. Unsupervised machine learning recognises underlying patterns in the input data and pre-processes data.
Deep learning is a computer model based on the neural networks of the brain to simulate brain processing and convolutional neural networks (CNN) is a commonly used algorithm to distinguish AMD from normal OCT images.
NLP is a form of AI which attempts to interpret human language in verbal and written text forms.
The relative performance of the three AI techniques depends on the selected algorithm, purpose of the study and input data set. AI techniques were adopted to study several diseases in articles considered in this review. Three studies used AI in AMD, including one study which adopted deep learning techniques and had an AUC 97% in differentiating AMD from normal OCT images.
With respect to DR, AI techniques were proven to be accurate indicators of DR and in some studies superior (81% AUC) compared to traditional indicators (69%) (Yoo and Park., 2013). These studies indicate that integrating AI techniques with EHR data is a promising model for improving early detection of patients living with diabetes at high risk of developing DR.
The evidence in this paper implicates AI technology as an accurate method for diagnosing and predicting disease progression and gives reason to expect that AI using EHR data will be applied more widely in ophthalmic care in the future.