Dr.Sabitha Krishnamoorthy, presented a talk on “Artificial Intelligence in Healthcare: Today and Tomorrow” at the 3rd American College of Physicians (ACP) India Chapter conference, on 2nd September 2018 at Lucknow, India.
“AI will not replace human physicians as long as we remain “humane Physicians”.
AI’s adoption in healthcare can be classified into 4 broad areas:
- Drug Discovery: AI has significantly made the new drug discovery process quicker and cheaper. With 90% of drug candidates failing to reach approval and the expense of clinical trial failure estimated to be up to $1.4 billion (over half of the average total new drug cost), pharma companies are embracing AI to make the process more cost efficient.
- Analysis and Prediction of Outcomes: The Permanente Medical Group’s division of research created a predictive analytic model to identify which hospitalized patients today are most likely to end up in the ICU tomorrow and could give alerts to notify physicians whenever a patient is deemed “at risk.” One of the largest Prospective cohort study (using the CPRD with routine clinical data of 378,256 patients from UK family practices who were free from cardiovascular disease at outset) concluded that the four machine-learning algorithms they included in the study were better than the established AHA/ACC risk prediction algorithm to predict first cardiovascular event over the next 10-years.
- Pattern Recognition in aiding diagnosis: Pattern recognition is one of the core aspects in the specialty of Radiology. A large number of AI companies got US FDA approved just in the last 1 year, majority of them involve radiological imaging. Deep learning algorithms have been able to diagnose the presence or absence of tuberculosis in chest x-ray images with astonishing accuracy. A deep learning algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs published in JAMA ,showed that the algorithm had 97.5% and 96.1% sensitivity and 93.4% and 93.9% specificity in the 2 validation sets.
- AI for Disease surveillance and hospital management: Automated surveillance systems have the capacity to detect patterns of disease outbreaks. For e.g. a Twitter-based supervised classification algorithm consistently predicted and mirrored the CDC data for influenza outbreak with 85% accuracy. John Hopkins Command Center was instituted to augment operational efficiency. Since the implementation of the command center in 2016, Johns Hopkins Medicine has seen many benefits such as a reduction of ED boarding by 20%, a reduction of OR holding by 80%, and 60% increase of patient transfers from other hospitals.
AI Tomorrow and the Future
AI will be in widespread use in various aspects of healthcare over the next 5 years. That said, there are a number of challenges to overcome. Most of hurdles involve the big data that the AI algorithms feed on. In healthcare, getting access to these datasets poses a wide range of issues right from patient privacy issues, to data accuracy, to quality assurance of AI programs.
- Ownership: Who owns these data? With the advent of regulations like GDPR, what if a patient requests deletion of patient data when such patient’s data has already been fed into an AI algorithm?
- Data accuracy: Apart from this, data is often subjective, inaccurate and unstructured. Consider the diversity of clinician’s notes in electronic medical records which makes it difficulty in interpretation. Such data is also scattered across multiple locations and clinical settings due to which capturing a full profile would be difficult. Also geographic boundaries and regulations prevent free movement of data.
- Quality of AI programs: Another major challenge is the assurance of quality of the various AI tools and models. These algorithms must be made available for peer review and must undergo verification of the key details such as the underlying algorithm code, and analysis. For example, the images on which the model is trained, the physicians with which it is compared, the features the neural network which were used to make decisions as a particular outcome. All this needs to be reviewed carefully.
The Future would see physicians practising with AI assistants in clinics. This would improve productivity and ability of physicians to delegate the “mundane” aspects of their practice to machines. Physicians of tomorrow will also benefit from acquiring basic technical know-how of AI systems so that the effect of AI can be synergised with the ability of human physicians to augment the delivery of health care.