• September 27, 2022
• By admin

Globalization of AI Requirements in Healthcare

WHAT IS MEDICAL ARTIFICIAL INTELLIGENCE?

The essence of evidence-based medicine is informing clinical decision making with insights from previous data. Statistical methods have traditionally approached this task by characterising patterns within data as mathematical equations; for example, linear regression suggests a ‘line of best fit.’ AI’s ‘machine learning’ (ML) techniques uncover complex associations that cannot be easily reduced to an equation. For example, neural networks, like the human brain, represent data through vast numbers of interconnected neurons. This enables ML systems to approach complex problem solving in the same way that a clinician would, by carefully weighing evidence and reaching reasoned conclusions. However, unlike a single clinician, these systems can observe and process an almost infinite number of inputs at the same time. For example, an AI-powered smartphone app can now triage 1.2 million people in North London to Accident & Emergency (A&E). Furthermore, these systems can learn from each incremental case and can be exposed to more cases in minutes than a clinician could see in a lifetime. This is why an AI-powered application can outperform dermatologists in correctly classifying suspicious skin lesions4 or why AI is being trusted with tasks where experts frequently disagree, such as detecting pulmonary tuberculosis on chest radiographs. Although AI is a broad field, this article focuses solely on ML techniques due to their widespread use in critical clinical applications.

WHAT ARE THE CURRENT TRENDS IN MEDICAL AI?

Aside from demonstrating superior efficacy, new medical technologies must also integrate with existing practises, obtain appropriate regulatory approval, and, perhaps most importantly, inspire medical staff and patients to invest in a new paradigm. These difficulties have resulted in a number of new trends in AI research and adoption.

AI excels at specific tasks:

The research has concentrated on tasks where AI can effectively demonstrate its performance in comparison to a human doctor. These tasks typically have well-defined inputs and a binary output that is easily validated. A digital photograph is used to classify suspicious skin lesions, and the output is a simple binary classification: benign or malignant. Under these conditions, researchers only needed to show that AI outperformed dermatologists in terms of sensitivity and specificity when classifying previously unseen photographs of biopsy-validated lesions.

AI is assisting doctors rather than replacing them:

Because machines lack human characteristics such as empathy and compassion, patients must believe that consultations are being led by human doctors. Furthermore, patients cannot be expected to trust AI; a technology shrouded in scepticism. As a result, AI is commonly used to handle tasks that are necessary but limited in scope enough to leave the primary responsibility of patient management to a human doctor. A clinical trial is currently underway to use AI to calculate target zones for head and neck radiotherapy more accurately and quickly than a human. Although an interventional radiologist is still ultimately responsible for delivering the therapy, AI plays an important background role in shielding the patient from potentially harmful radiation.

AI assists under-resourced services:

Because a single AI system can support a large population, it is well suited to situations where human expertise is scarce. There is a shortage of radiological expertise at remote centres in many TB-endemic countries. Radiographs from these centres could be interpreted by a single central system using AI; a recent study found that AI correctly diagnoses pulmonary tuberculosis with a sensitivity of 95% and specificity of 100%. Furthermore, under-resourced tasks with unsatisfactory waiting times for patients are appealing to AI in the form of triage systems.

AI is a picky eater:

Developing ML models necessitates the collection of well-structured training data about a phenomenon that is relatively stable over time. A deviation from this results in ‘over-fitting,’ in which AI places undue emphasis on spurious correlations within previous data. Google attempted to predict the seasonal prevalence of influenza in 2008 using only search terms entered into its search engine. Because people’s search habits change so dramatically with each passing year, the model was so inaccurate in predicting the future that it was quickly abandoned. Furthermore, data that has been anonymized and digitised at the source is preferable because it aids in research and development.

POSSIBILITIES FOR THE FUTURE IN GENERAL PRACTICE:

AI will extract critical data from a patient’s electronic footprint. This will initially save time and increase efficiency, but after adequate testing, it will also directly guide patient management. Consider a consultation with a type 2 diabetes patient; currently, a clinician spends a significant amount of time reading outpatient letters, checking blood tests, and locating clinical guidelines from a number of disconnected systems. In contrast, given the patient’s clinical record, AI could automatically prepare the most important risks and actions. It could also automatically convert the recorded consultation dialogue into a summary letter for the clinician to approve or amend. Because they assist rather than replace clinicians, both of these applications would save significant time and could be implemented quickly.

As these systems’ validation improves, they will be given more responsibility. Instead of a rigidly defined ‘one-size-fits-all’ algorithm, AI could determine the threshold of statin initiation for the patient with type 2 diabetes on an individualised basis based on the patient’s history. The research required for this ‘personalised’ medicine would only be possible if AI intelligently summarised vast amounts of medical data. Furthermore, because AI can monitor millions of inputs at the same time, it will play a significant role in preventative medicine. When AI determines that a patient’s risk of developing a specific diabetic complication warrants intervention, it could suggest consultations proactively. In contrast, assigning a human to the task of closely monitoring every test would be impractical.

Artificial intelligence-powered systems will also bring specialised diagnostic expertise into primary care. If an image of a skin lesion is sufficient to accurately diagnose its cause, it could be captured at a GP practise and sent to a dermatology AI system for immediate analysis. Patients identified as low risk would receive immediate reassurance, whereas high-risk patients would have shorter referral wait times because clinics would only receive selected cases. This concept is not limited to skin lesions; AI has demonstrated the ability to interpret a wide range of image data, including retinal scans, radiographs, and ultrasound. Many of these images are easily captured using inexpensive and widely available equipment.

Future AI research should focus on carefully chosen tasks that broadly align with the trends described in this article. Integrating these systems into clinical practise necessitates the development of a mutually beneficial relationship between AI and clinicians, in which AI provides clinicians with increased efficiency or cost-effectiveness and clinicians provide AI with the necessary clinical exposure to learn complex clinical case management. Throughout the process, it will be critical to ensure that AI does not obscure the human face of medicine, because the public’s reluctance to embrace an increasingly controversial technology will be the biggest impediment to AI’s widespread adoption.