Artificial intelligence and Machine learning in Healthcare

Karthikeyan Anantha Subramanian
Monday, 27 August 2018
Dais the next oil…the key lies somewhere in the historic data..

Technological innovations are penetrating the healthcare industry across the world today, with the focus shifting towards prevention rather than cure, the proliferation of Digital Health tools, including mobile health apps and wearable sensors, holds great promise for improving human health. Thousands of health apps are now available on top app stores worldwide with more than 200 health apps being added each day.

While majority of mobile health apps available focus towards general wellness, the number of apps for management of specific health conditions – often associated with patient care are increasing at a fast pace. These companies are turning towards AI for a variety of applications, including diagnostics, prevention, clinical workflow and even fraud detection in claims

Wearables

Wearables play a significant role, most convenient tools to collect health data, monitor and interact with users on the go. Helping people lead healthier lifestyles and achieve sports goals is one thing, saving people’s lives is an entirely new level, and powered by AI. Wearables can be an extremely useful tool for people with chronic health conditions related to heart

Wearable EKG/ECG monitors come in various shapes and sizes today, that can be worn on the wrist, chest and even embedded into clothing. These can help get medically accurate EKG readings outside of a medical setting.

Interpretation of ECG Signals

Noise Reduction: ECG signals are affected by noise due to several factors, including positioning of the device, interference from other devices, power interference etc. The first step in ECG signal processing is noise reduction and removal of baseline wander using digital signal filters and other relevant algorithms.

Feature Extraction: Extracting relevant features including intervals like QRS Complex, ST Segment, PR Segments, RR Intervals, amplitude of various waves and the variances over a period of time, a combination of wavelet transforms and algorithms like Pan Tomkins algorithm is used for this purpose, while this seems straightforward necessary care has to be taken while handling signals from different sensors for interoperability and in case of missing waveforms

Analysis: A variety of approaches could be taken towards interpretation of the signals post all the pre-processing and feature extraction, several algorithms are even FDA approved owing to higher degree of accuracy, some of the possible options include

  • Statistical methods and algorithms to determine the intervals and relation between various features over a period of time to determine potential conditions
  • Standard supervised models using support vector machines or ensemble models like random forests can provide a high degree of accuracy if modeled right
  • Deep learning methods using Convolutional and Recurrent Neural Networks for detection of anomalies

we at Cover2Protect have taken a hybrid approach towards analysis signals and use a combination of above methods based on factors like the type of signal – 1 lead, 2 lead, quality and duration of signal, historical data availability for an individual and the conditions in focus as some methods are better suited for detection of certain conditions. Our team of data scientists are also working on image processing algorithms that can identify a set of conditions

Key to achieving required results depends on a large and properly annotated high-quality data set and choosing the algorithms best suited