Source: Deep Learning on Medium
Everyday, we hear about artificial intelligence (AI) algorithms beating radiologists, pathologists and other doctors in their respective fields. Companies ranging from Google to one man startups are building AI algorithms that can predict, diagnose and treat diseases. Large health systems are investing heavily in AI and professional associations like Radiological Society of North America is hosting competitions in AI. NEJM, JAMA and other major medical journals are publishing studies in medical AI. In the near future, we all will be using AI algorithms knowingly or unknowingly while caring for our patients. Hence we need to understand how medical AI algorithms are made, potential use cases, pitfalls and ethical implications. This article is the first in a series of articles discussing these topics. AI, its subsets and medical use cases are discussed in this article.
Artificial intelligence and its subsets
The term artificial intelligence was said to be first coined by John McCarthy in 1956. In general terms, artificial intelligence is any non living entity that imitates human intelligence and can perform tasks in which humans use their intelligence. Machine learning, a subset of AI is defined by Arthur Samuel in 1959 as a field of study that gives computers the ability to learn without being explicitly programmed. In traditional computer programming, a computer programmer gives specific instructions for the computer to conduct a specific task. In machine learning, the computer is able to figure out the best solution for a problem. At present most of the activities and hype is happening inside the machine learning sphere. If we have data on different factors like family history of coronary artery disease, presence of high blood pressure, diabetes etc on a wide variety of patients, we can use this data to create a machine learning model to predict who will develop coronary artery in the future. Recently, researchers at the Mayo clinic created an algorithm that could predict acute kidney injury in ICU patients 6 hrs before in happens. Deep learning, a subset of machine learning has taken the spotlight in solving problems like medical image diagnosis.
In the above example of predicting coronary artery disease, we fed the algorithm with factors that are associated with coronary artery disease. But we cannot define all the factors in a 3 dimensional CT scan image that makes a lung nodule cancerous or define all the features that makes a mammogram malignant. Given enough data, deep learning can identify these underlying factors and predict an outcome. Deep learning does this by mimicking human neuronal circuitry, using neural networks. Neural networks consists of multiple nodes. Nodes are places where computation happens. Each circle in the picture below represents a node, which is analogous to a neuron in our brain. As you can see in the figure below, these nodes are interconnected. When you stack multiple neural networks, you get a deep learning network.
Representation of Deep Learning architecture
Applications of AI in healthcare
Medical image classification
One of the most active areas of AI research in healthcare is medical image classification. There are a few FDA approved applications already in the market. Aidoc has FDA approved software that can detect pulmonary embolism, pneumothorax, rib fractures, lung nodules and intracranial hemorrhage. Their workflow optimization software flags life threatening disease so that radiologist can prioritize those studies. IDX-DR uses AI to analyze retinal images to detect diabetic retinopathy. This could be used by primary care physicians in their offices or can be deployed in remote areas without tertiary care centers. Viz.AI optimizes stroke management by identifying large vessel occlusions and alerts the on call stroke team autonomously. Zebra medical vision has an AI medical image analysis platform that has coronary calcium scoring algorithm, chest x-ray triage algorithm etc. Automated polyp detection during colonoscopy is another application of AI in medical image diagnosis.
Alpha numeric data
Electronic medical records generates petabytes of data. This is one of the most underutilized resources in healthcare, with medical AI this is changing. Machine learning algorithms in this arena includes acute kidney injury prediction algorithm, mortality prediction algorithm, diabetes detection algorithm etc.
Medtronics currently has algorithms that can predict low blood glucose and suspend insulin pump. Their automated insulin pump has been in the US market for sometime. Tandem, Omnipod and Tidepool has algorithms that helps manage diabetes.
Demographics data has been combined with EMR data to predict patient no shows. This helps in optimization of patient scheduling there by increasing access to healthcare. Empatica uses data from wrist band to predict seizures. There are multiple companies involved in the analysis of electrocardiograms. Afirma uses machine learning to comb through molecular markers to predict whether a thyroid nodule is benign or malignant.
AI is also currently being used for drug discovery. This approach promises to reduce the cost of drug discovery. Sanofi recently partnered with Google to expedite drug discovery. Cloud pharmaceuticals uses existing toxicity data to predict whether a new molecule is toxic or not.
Natural language processing (NLP)
NLP uses AI to understand and manipulate human generated speech or text. NLP can be used to extract information from clinical notes and imaging reports. This can be used to generate disease specific tags and can help in coding and billing. 3M’s CodeRyte CodeAssist System can autonomously generate CPT and ICD codes from clinical documents. NLP can also be used to generate automated summaries for individual patients. Amazon’s Comprehend Medical helps to combine medical data from different sources. Information generated by this service could be used for clinical trial management, medical decision support and revenue cycle management. NLP can also be used to find at risk patients who will benefit from early preventive interventions. Another use case is autonomous generation of documentation for prior authorization. NLP could also be used to develop virtual scribes. Nuance’s Ambient Clinical Intelligence technology promises to free physicians from the burden of clinical documentation and increase physician patient communication and interaction.
This article barely scratched the surface of medical AI. There are many more companies and educational institutions working on healthcare related AI and they are coming out with new products everyday.
Keep yourself up to date on medical AI at