SHB: Computational Algorithms for Predictive Health Assessment

Prinicpal Investigator:

Co-Investigators: , ,


Project Summary

The main goal of this work was to develop algorithms for early illness recognition in elderly. Early illness recognition (EIR) is important, as research has shown that results in better medical outcomes and a reduction in health care cost. We developed methodologies (see Figure 1) that link sensor data to the medical (nursing) records for monitoring the residents of TigerPlace, an aging in place community from Columbia, Missouri. The developed methodologies addressed several of the EIR challenges: lack of training data, ground truth uncertainty and processing multiple time series. To address the lack of training data we investigated an abnormal event detection methodology based on one-class classifiers (OCC). While OCC are not as good as regular classifiers, they can be successfully employed for anomaly detection. To account for the ground truth uncertainty, we developed a multiple instance learning methodology which allows linking of loosely coupled sensor and EHR patterns. In this approach we only need to know that the resident had some problems during a given day without requiring the knowledge of the exact time of the incident. Our most interesting methodology developed in this project was converting the sensor data to a discrete format similar to a genomic sequence. This approach enabled the utilization of known bioinformatics algorithms such as Smith -Waterman or motif discovery for EIR. Under this approach, frequent temporal sequence patterns were extracted daily from the resident data. When a significant number of them were missing in a given day, the algorithm would infer that something was wrong with the resident. Comparing sensor sequence patterns between multiple residents using multiple sequence alignment, will enable researchers to observe similar behavior elements and develop generic EIR algorithms that would not require individualized training.

Early Illness Recognition Methodologies


Z. Hajihashemi, M. Popescu, “A Multidimensional Time-Series Similarity Measure With Applications to Eldercare Monitoring,” IEEE Journal of biomedical and health informatics, 2015, vol. 20, no. 3, pp. 953-962.

Hajihashemi Z & Popescu M, “A New Illness Recognition Framework Using Frequent Temporal Pattern Mining,” Proceedings, ACM UbiComp International Workshop on Smart Health Systems and Applications, September 13-17, 2014, Seattle, WA, pp 1241-1247.

Hajihashemi Z, Yefimova M & Popescu M, “Detecting Daily Routines of Older Adults Using Sensor Time Series Clustering,” Proceedings, IEEE International Conference of the Engineering in Medicine and Biology Society, Chicago, IL, August 26-30, 2014, pp 5912-5915.

M. Yefimova, T. Banerjee, Z. Hajihashemi, DL Woods, M. Popescu, M. Skubic, M. Rantz, J. Keller, “Characterizing Trajectories of Daily Routines of Older Adults with Sensor Technology”, 47th Annual Communicating Nursing Research Conference, Seattle, WA, April 9 – 12, 2014.

Hajihashemi Z & Popescu M, “An Early Illness Recognition Framework Using a Temporal Smith Waterman Algorithm and NLP,” Proceedings, American Medical Informatics Association (AMIA) Annual Symposium, Washington DC, Nov. 16-19, 2013.

Hajihashemi Z & Popescu M, “Detection of Abnormal Sensor Patterns in Eldercare,” E-Health and Bioengineering Conference (EHB), Iasi, Romania, November 21-23, 2013, pp 1-4.

Hajihashemi Z & Popescu M, “Predicting Health Patterns Using Sensor Sequence Similarity and NLP,” Proceedings, IEEE International Conference on Bioinformatics and Biomedicine (BIBM), Philadelphia, PA, October 4-7, 2012, pp 948-950.

Mahnot, M. Popescu, “FUMIL-Fuzzy Multiple Instance Learning for Early Illness Recognition in Older Adults”, IEEE WCCI, Brisbane, Australia, 2012, pp. 2102-2106.Z. Hajihashemi, M. Popescu, Predicting Health Patterns Using Sensor Sequence Similarity and NLP, Proc. of BIBM 2012, Philadelphia, PA, 2012, pp 948-950.

M. Popescu, A. Mahnot, “Early Illness Recognition in Older Adults Using In-Home Monitoring Sensors and Multiple Instance Learning”, Methods of Informatics in Medicine, 2012 Aug 7;51(4):359-67.