Artificial Intelligence: Highlights in Diagnostics

Artificial Intelligence (AI) has been a disruptive technology in numerous fields of industry – and the same is expected to happen in the life sciences, with advancements in machine learning and neural networks already seeing applications. Much of the pharmaceutical and healthcare industry depends on computational power, such as drug discovery, bioinformatics, and in silico trials. However, even areas which have not been traditionally dependent on computational power, will be expected to see advancements. One of these areas is diagnostics, particularly for complex health conditions – where AI can lead to breakthroughs by bringing about diagnoses through more detailed, timelier and deeper examinations. We will discuss examples of where AI has already been deployed – and then outline the common challenges faced by all these areas for future progression in its use.

Practical Highlights

Using machine learning through deep neural networks for image recognition can provide valuable tools for physicians. With AI trained to detect disease, multiple advances have already been realised. In 2017, groundbreaking research on skin cancer compared the performance of convolutional neural networks to board certified dermatologists in diagnosing cancer. The network required no hand-crafted procedures and was able to work with both photographic images and images acquired through biopsies or dermoscopies – something that previously required significant preprocessing. The network was able to match the performance of at least 21 dermatologists in classifying keratinocyte carcinomas and melanomas. It also shows the potential to be deployed through mobile devices, posing revolutionary possibilities for the future of diagnosing and differentiating between malignant and benign tumors.

A 2020 study additionally found it possible to train AIs with a unique neural network to draw conclusions from chest radiographies. The performance of the AI was found to often be comparable to third-year radiology residents, although human input was sometimes needed to ensure accurate diagnostics; the researchers noted that this was more associated with the necessity of yet more diverse training images for the network rather than any other shortcoming, highlighting the data-hungry nature of AI applications.

Another area where complex diagnostics are required is the field of haematological disorders. Artificial Intelligence has been employed through multiple methods such as:

• Cytomorphology – where the AI makes assessments based on images of cells, aided by microscopy
• Cytogenetics – with AI performing examinations on the structure and banding of the chromosomes of a patient
• Immunophenotyping – where fluorescent antibodies are targeted towards sites of interest – the tumour – through diagnostically relevant antigens; AI can be used to perform a diagnosis based on the light emission profiles presented by antibodies
• Genetics – With high-throughput sequencing, it is possible to use machine learning algorithms to perform predictive analyses of clinically relevant genes; this approach can also be used to predict treatment responses, much like Digital Twins, which are discussed below

These advances are not limited to haematology, and show the flexibility of artificial intelligence across a number of areas – with enough training data, the possibilities are limitless. One challenge that is common across all discussed methods is the need to minimize human involvement – which may not always be achievable through raw increases in computational power, but may require more creative solutions instead. 

A more contemporaneous application of Artificial Intelligence has been in the diagnosis of COVID-19 infection using chest Computed Tomography scans. SARS-CoV-2 can be reliably identified by PCR, but given the shortage of PCR test kits earlier in the pandemic, AI was able to demonstrate equal competence to a senior thoracic radiologist. Surprisingly, the convolutional network model even detected 68% of patients who had been classed as COVID negative by radiologists, but were proven to be positive using PCR. Additionally, commercialised firms such as diagnostics.ai and pcr.ai have already begun bringing technology to automate the PCR infection diagnosis process to the market.

Google Health has also demonstrated significant advances in bringing end-user operated diagnostics – by enabling patients to take their own pictures, which are then evaluated by an AI. Models employed by Google Health have demonstrated performances that exceed or match clinical operators in diagnosing retinal health conditions. Another Deep Learning System developed by them has also shown potential in differentiating between 26 skin conditions, representing the vast majority of reported skin disease – with the system having shown competence comparable to nurses and primary care physicians.

A particularly promising application of AI are Digital twins. This is a concept originally developed to ascertain system failure – particularly in mechanical engineering. However, AI promises to translate the effects of digital twins to the field of diagnostics. The technology is centered around constructing models based on all recorded variables of a system, which in this case would be biomarkers from a patient as well as environmental factors relevant to the disease in question. These models can then be analysed for the effects different variables have on them. This is not limited to diagnostics however –  models can also be treated with all possible drug candidates to ascertain their effects. Naturally, such modelling requires gargantuan data sets not only on biomarkers, but also deep understandings on the Pharmacokinetic and Pharmacodynamic workings of candidate products – as well as their exact mechanisms of action. 

Avenues of Progress

The limitations are eminently visible in current AI technology. The biggest predictor of performance across all areas has been the availability of training materials – often running up to hundreds of thousands of images required to acclimatize models to their tasks. Producing the volumes of such data will require massive advancements in computing. Cloud and quantum computing aim to alleviate such obstacles. The ability of quantum computing to process vast amounts of information at unprecedented speeds is expected to be incredibly synergistic to the rise of Artificial Intelligence. 

But we can also minimize the need for more datasets by enacting more creative statistical solutions. Other industries have already exhibited the data-minimizing effects of employing techniques such as nearest-neighbour modelling. More stringent feature selection criteria can also lead to more data-efficient AI models in the short-term. This is a wide theme in the industry – as Big Data has become trending, we need to re-examine our criteria for selecting which parts of the dataset truly deserve to be included in training – as was mentioned in the last Information Commissioner’s Office report. 

The future looks bright for Artificial Intelligence – and the avenues that it could explore in diagnostics, both in imaging and in predictive analyses, are nearly limitless. The barriers remain mostly technological – although data management practices and model design can provide immediate benefits even before computational power increases. Join us in our Bioinformatics Strategy Meetings, where leading experts from the industry will discuss the significance and the future of AI, and other computational methods, in the life sciences!