Medical Device Data Analytics
The volume of data generated by portable devices along with the richness of available algorithms, enable machines to learn complicated patterns such as the relationship between the measured signal and the true physiological state, or degradation in the sensor’s accuracy. Currently, bio-sensors with RF modules and smartphones can help monitor bio-signals from a distance, but actions, like calibration and drug administration, are still done by humans. Advanced analytics can minimize human involvement, and take disease management a step forward toward complete automation. Combined with other data modalities it can become a powerful resource for driving better patient management and disease insights.
Our team is developing machine learning and hierarchical dynamic models to analyze medical device signals for different use cases. We combine deep learning algorithms such as feed-forward and recurrent neural networks, with classic machine learning algorithms (logistic regression, random forest, gradient boosting trees) to analyze signals reported by sensors. We also develop pipelines that combine modules for multiple data modalities to predict the true physiological states from bio-sensors output, and model sensors drift, or disease progression.
Our recent work involves the analysis of continuous glucose monitoring (CGM) data from cohorts of diabetes patients, analysis of multi-modal electrocardiograms combining EEG, EOG, EMG and heart rate for analyzing sleep patterns and disorders. In our current efforts, we combine other modalities, like the clinical history of the patients, into one ensembled pipeline. Such pipelines can then be used for personalized decision-support and better patient disease management. They can also help us better understand disease progression and onset, and identify the right points in time for preventive interventions.