In healthcare, patients are diagnosed based on symptoms, the results of certain diagnostic tests, historic data, and other factors. Based on this information, a physician will assign the patient a corresponding treatment.
There are situations where a diagnose is dependent on thousands of different data points.
Deep learning algorithms can be applied to analyze large corpora of patient phenotypes, providing a more data-driven approach to patient categorization.
Naturally, there is no one algorithm that can flawlessly diagnose any disease. Diagnostic assistance is used for assisting the physicians in real time by either data retrieval or diagnosis recommendations. There are also several companies providing a prediction service for identifying most at-risk patients.
Just as Google’s machine learning algorithms can identify specific objects and people from images, the AI-powered image recognition can also be used to identify and classify medical images (e.g. to recognize lesions and nodules; localize organs, regions, landmarks, etc.). Such discoveries can be assessed by human experts to reach more accurate diagnoses.
AI-enabled image recognition is applicable to finding diabetes by looking at retinal images or detecting anomalies in heart activity, and many other areas where there’s a need to detect specific objects/diseases/anomalies.
In medical imaging, there are already a handful of highly accurate classifiers that are comparable to clinician performance. For a comprehensive overview of deep learning achievements in this field, we recommend that you read this recent study.
In radiology for example, where the task of the physician is to diagnose a patient using medical imaging, computers have been taught to identify pathologies from such images at either a comparable or, in some cases, even higher level than human doctors.
Left: Images from two lymph node biopsies. Middle: earlier results of our deep learning tumor detection. Right: Google’s current results – Image source
AI-enabled text mining in healthcare
In line with the rapid growth of scholarly publications and electronic health records (EHRs), there’s an increasing potential to apply machine learning to biomedical text mining.
The principal tasks in medical text mining include named entity recognition, relation/event extraction, and information retrieval.
Deep learning applications, tasks, and models based on NLP perspectives. – Image source
In December 2016, IBM announced that Pfizer would be one of the first organizations to utilize the Watson for Drug Discovery cloud-based platform. To provide context, the deep learning program has accumulated data from 25 million Medline article abstracts and 1 million medical journal articles (a human researcher can read 200-300 articles in a year).
The potential use cases for text analysis in biomedicine include:
- Generating a knowledge base by having a machine “read” all publications as they are released.
- Translating a free-form anamnesis into machine-readable data about the patient’s treatment history.
- Understanding patients’ description of their health issues.