This project leverages ongoing work at Tiger Place (TP), University of Missouri (MU) in the use of sensor technology for in-home health assessment. The TP team has deployed sensor networks in the homes of seniors, with a wide range of sensor types and analysis approaches. They are integrating their sensor networks with an in-house nursing electronic health record (EHR) and investigating health context-aware computational algorithms for health and wellbeing assessments. The proposed project has the following objectives: (1) Integrate the sensor network with an EHR developed in-house to provide automatic health context for comprehensive algorithm development; (2) Investigate algorithms for identifying health patterns based on sensor data and contextual health information such as chronic conditions and medication changes that are provided by the EHR data; and (3) Investigate the possibility of predicting physiological changes such as blood pressure based on sensor data. A variety of machine learning methods are investigated for predictive health assessment. There are two potential difficulties that this research tackles: ground truth uncertainty and data unbalance. To address these problems the project is developing two new machine learning methods: a fuzzy extension of multiple instance learning and a sensor firing sequence similarity based method for recognizing pattern changes. Existing sensor data is used as a starting point to develop the proposed methods, along with simulated data for more diverse testing scenarios. The integrated combination of the sensor network and EHR is expected provide a unique, rich dataset in which to investigate health context-aware algorithms.
Many older adults in the US prefer to live independently for as long as they are able, despite the onset of conditions such as frailty and dementia. Elderly patients are particularly at-risk for late assessment of health changes due to factors such as their impression that such changes are simply a normal part of aging or their reluctance to admit to a problem. In-home sensors networks have emerged in the last ten years as a possible solution for early illness recognition. Many projects such as CASAS, TigerPlace and ORCATECH have demonstrated the utility of in-home sensors for monitoring elderly but also have shown the necessity of developing new pattern recognition algorithms able to handle large amounts of diverse data also known as big data. In TigerPlace, an aging in place facility from Columbia, MO, we created a unique living laboratory by deploying in-home sensors together with electronic health record (EHR) system developed in-house. In-home monitoring devices such as infrared motion detectors, Kinect depth cameras, Doppler radars and bed sensors capture information related to the behavior of the residents from the monitored apartment and assist the clinical personnel in medical decision making. The goal of this NSF project is to investigate early illness algorithms that link the data provided by the in-home data to the clinical data from the EHR. There are two main problems that were addressed this endeavor. First, the mapping between the EHR data and sensor data has a high degree of uncertainty. The main reason for this temporal uncertainty is the delay between the onset of the health condition that produced the behavior captured by the deployed sensors and the related data entry in the EHR system made by the nursing personnel. To address this problem, we investigated a multiple instance learning approach (MIL), in which classifiers are trained with sets of data at a time to label a day as normal or abnormal based on sensors and EHR data for a given resident. In an experiment performed on data from six TigerPlace residents, a MIL classifier was about 10% superior to a traditional anomaly detection approach. We further investigated a fuzzy extension of the MIL framework (FUMIL) which, on a pilot dataset that consisted of three TigerPlace residents, outperformed the traditional MIL by about 10%. Second, the large amount of data and its diversity (big data) made it difficult for clinicians to interpret sensor patterns. To address this problem we employed an approach used in bioinformatics to interpret new genomic sequences based on "guilt by association": if they are similar, they should (might) have the same function. In our case, this translated into: if the sensor sequences are similar, they should be due to the same behavior and, consequently, might have the same clinical meaning. We note that the assumption that a given clinical condition maps to a given set of behaviors, might not hold for a diverse population, but we found it true in elderly. The clinical meaning of each daily sensor sequence was obtained by mapping EHR nursing notes to the National Library of Medicine (NLM) Unified Medical Language System (UMLS) concepts using natural language processing (NLP). A daily sequence was obtained by concatenating all sensor firings mapped to set of symbols together with their time stamps. Since sensor sequences, unlike the genomic one, have a temporal dimension, we proposed a new temporal version of the well-known bioinformatics similarity measure, Smith-Waterman (SW), denoted as temporal SW (TSW). On a pilot dataset obtained from three Tiger Place residents, with a total of 1685 sensor days and 626 nursing records, we were able to annotate daily sensor sequences with UMLS concepts with an average precision of 0.64 and a recall of 0.37. In this project we demonstrated the utility of combining sensor and EHR data for early illness recognition. The project has also resulted in 1 journal and 7 conference publications.