Artificial Intelligence in Medicine: A Snapshot of Current Trends
Artificial intelligence (AI) is set to transform healthcare, and its use is becoming a reality in many medical fields and specialties. As early as 2016, the healthcare sector attracted more investments in AI projects than any other sector of the global economy. While these investments totaled $6 billion in 2013, they reached $170 billion in 2021, or 30 times more when adjusted for inflation .
Evidence-based medicine is about making informed clinical decisions based on patterns and insights derived from historical data. Traditionally, statistical methods have been used to characterize patterns in datasets using mathematical equations. With 'machine learning', AI now offers techniques capable of uncovering complex associations by analyzing vast amounts of data stored in the form of digital images or electronic patient health records, hardly reducible to equations.
AI and machine learning can significantly improve the quality of patient care by increasing diagnostic accuracy and thus clinical outcomes. At the same time, AI-powered tools also make healthcare more affordable and accessible, thereby reducing inequalities in research and healthcare delivery. The technology carries obvious uncertainties and risks; it must integrate with current medical practices, receive appropriate regulatory approval and, ultimately, inspire trust among doctors and patients to increase the rate of AI adoption.
What are the most common uses of AI in medicine today?
1) Early disease detection and diagnosis – aid in precision diagnosis
As early as 1959, Keeve Brodman and colleagues claimed that “the making of correct diagnostic interpretations of symptoms can be a process in all aspects logical and so completely defined that it can be performed by a machine.” Years later, William B. Schwartz predicted that by the year 2000, computers would play an entirely new role in medicine, acting as a powerful extension of the physician’s intellect.
Machine learning is capable of recognizing patterns just as doctors see them. The key difference is that algorithms need to see several thousand concrete examples to achieve this. With this in mind, the system can rapidly process an infinite amount of digital patient data in a fraction of a second, spotting and processing far more cases than a physician could see in several lifetimes.
The use of AI and machine learning has already entered mainstream medical practice to interpret certain health conditions based on a variety of numerical datasets and digital images, such as:
· Detecting lung cancer or strokes based on CT scans
· Assessing the risk of sudden cardiac death or other heart diseases based on computer reading of electrocardiograms and cardiac MRI images
· Classifying skin lesions as benign or malignant using digital skin photographs
· Finding indicators of diabetic retinopathy in retinal photographs
AI and machine learning have proven their value in helping the healthcare provider spot unusual or atypical features on diagnostic images, thereby reducing the number of errors. But this is just the beginning: more ambitious systems will involve linking multiple data sources (CT, MRI, genomics and proteomics, patient data, laboratory data, even handwritten files) as well as improving their interoperability to evaluate pathologies or their progression.
2) Acceleration of drug development
Developing drugs is a notoriously costly endeavor. Many of the analytical processes involved in drug development can be made more efficient using machine learning. This has the potential to save years of work and hundreds of millions in investments.
AI is already successfully used in the 4 main stages of drug development:
· Stage 1: Identifying targets for intervention: High throughput techniques such as short hairpin RNA (shRNA) screening and deep sequencing have greatly improved the discovery of viable target pathways. However, integrating numerous and varied data sources – and identifying relevant patterns – remains a challenge for traditional techniques. Machine learning algorithms simplify the processing of available data and enable the automatic identification of relevant target proteins.
· Stage 2: Discovering drug candidates: Currently available software is often inaccurate at suggesting drug candidates, and it can prove time-consuming to shorten the list of lead candidates. In contrast, machine learning algorithms can predict the relevance of a compound from its structural fingerprints or molecular descriptors. They then screen several millions of potential candidates and filter them down to the best options, with the least side-effects. This saves considerable time in drug design.
· Stage 3: Speeding up clinical trials: AI and machine learning can improve, simplify and accelerate clinical trials. The system can optimize patient selection and recruitment. It can generate synthetic data to complement insufficient datasets. For example, it can generate a synthetic control group by matching the clinical characteristics of study participants with patient ‘outcome profiles’ stored in a database for a given disease. This would enable the use of synthetic patients for simulating real-time diagnostic findings, that could be linked to likely therapeutic outcomes. AI and machine learning techniques could also be used to better predict and understand potential adverse effects, and to identify patient sub-populations. However, the use of this technology introduces a set of uncertainties that must be dealt with both in clinical trial protocols and reports.
· Stage 4: Finding Biomarkers to diagnose a disease: Discovering suitable biomarkers, whether diagnostic, predictive, risk or prognostic …for a particular disease is hard. It’s another expensive time-consuming process that involves screening tens of thousands of potential molecule candidates. AI and machine learning can assist with biomarker identification through automation of much of the manual work and speed up the process. These algorithms classify molecules into good and bad biomarkers, enabling clinicians to focus on the best prospects.
3) Personalize treatment
Patients differ in their response to drugs and treatment schedules. Personalized treatment therefore offers enormous potential for improved patient outcomes.
The machine learning system can automate this complex statistical work and reveal the characteristics that predict a patient's chances of responding to a given treatment. To do this, the system cross-references similar patients and compares their treatments and outcomes. Predicting the outcome considerably simplifies the implementation of appropriate treatments at lower costs.
4) Improve gene editing
A major goal of gene editing is to remove harmful genes that cause disease. Although traditional technologies such as CRISPR have made great strides, the risk of error remains significant. Machine learning algorithms can help identify the precise location for editing and ensure the correct replacement of the DNA strand, thus reducing the risk of error throughout the process. The new algorithms can predict the effectiveness of different genome-editing-repair options. Unintentional errors in the correction of DNA mutations in genetic diseases can thus be reduced.
In traditional clinical research, when progress leads to the discovery of a new drug for a defined condition, the criteria for evaluating and approving a drug as a new therapeutic agent are well established.
When the intervention is an AI-powered tool rather than a drug, the medical community expects the same level of safety and efficacy, but the standards for describing and testing AI and machine learning interventions are far from clear. Numerous algorithms rely on highly complex and difficult to decrypt mathematics, sometimes referred to as the ‘black box’, to get from the input data to the final results. An inability to ‘unpack the black box’ and clarify the inner workings of an algorithm (e.g., how a specific set of data leads to a prognosis) might have an impact on the probability that the FDA will approve a dossier that relies on an AI-based trial. One of the key concerns for regulators is to ensure that AI and machine learning applications work as advertised in multiple-use settings.
The US FDA has approved some assistive algorithms, but no universal approval guidelines currently exist.
AI adoption in healthcare continues to face challenges, such as the lack of trust in the results delivered by a machine learning system and the need to meet specified standards. Can AI and machine-driven care meet the standards we demand of a new therapeutic intervention? Algorithm outputs are expected to be applicable across a wide range of patient populations and disease prevalences.
Although AI and machine learning applications carry obvious uncertainties and risks, they also offer the potential to significantly improve the quality of patient care and clinical outcomes, while making care more accessible and affordable for patients.
But this is just the beginning, as the future promises even greater improvements. The more we digitize and integrate our medical data, the more we will be able to use AI to help us find useful, actionable patterns – patterns we can use to make more accurate, cost-effective decisions.