Data, Deep Learning and (medical) Diagnostics
Radhika Iyengar, Solution Head – Diaspark Healthcare
Certain presentations of common diseases and unusual cases in Clinical pathology can pose diagnostic challenges even to the expert physicians.
The clinical features of certain clinical cases cannot immediately equip a doctor to arrive at accurate diagnosis. Providers need to scour though multiple medical records to be able to characterize disease profiles.
Recent advances in machine learning could potentially make it easier for doctors to sort through medical information that are stored as images (MRIs..), unstructured data (providers notes) and structured records (laboratory test results).
Since all this is data, medical diagnostics is basically a data problem. The various electronic medical records contain valuable information and learnable patterns. Recent breakthroughs in neural networks (deep learning algorithms) have enabled computers to find fine patterns and turn huge data of this kind into deep insights. Such technologies are exploited to understand unstructured data which constitute 80% of the EMRs and make that information actionable.
This implies that we could leverage such machine-learning technology to transform diagnostic healthcare.
As physicians are inundated with diagnostic data, deep-learning algorithms can be used to teach machines to identify patterns. Meaning – tools could be created to mark medical images requiring further examination by doctors, thereby saving them the effort and time of examining numerous images. Doctors could also be enabled with such tools to scan the EHRs for any specific condition.
For example – a condition like sleep apnea can be inferred even if the patient medical record does not explicitly state ‘sleep apnea’. Text analysis incorporating deep learning can infer from the facts about the patient – heavy snorer, is obese, having low blood oxygen levels and is taking Provigil – that the patient likely has ‘sleep apnea’.
Hospitals for instance can use, if not already, such data analytics to identify cardiac patients at greatest risk while reducing the number of re-admissions. Such initiatives in clinical analytics while currently tied to cost-saving and resource-conserving efforts, could also lead to improved population health.
Physicians currently lack the tools needed to derive valuable insights from the diverse forms of medical data, particularly automation tools that employ pattern recognition algorithms to, for instance, characterize disease profiles. Some believe that data-driven technologies like deep learning could replace 80% of what providers do.
This provides an opportunity for health IT product vendors and service providers to expand their offerings in health analytics via engaging machine-learning specialists, data scientists and GPU programmers.