Second Workshop on Artificial Intelligence for Function, Disability, and Health

Functional Status Information (FSI) describes physical and mental wellness at the whole-person level, and includes information on activity performance, social role participation, and environmental and personal factors that affect well-being and quality of life. Collecting and analyzing this information is critical to addressing the data needs in caring for aging global populations, and providing effective care for individuals with chronic conditions, multi-morbidity, and disability. In this paper, we present a first step towards the design of AI-enabled system increasing the self-awareness of own functional status. In particular, we focus on the underlying semantic layer of the such an AI-enabled system by presenting two modules extending the HeLiS ontology aiming (i) to represent barriers in improving own functional status and (ii) to support the conceptual representation of the arguments leading to the detection of such barriers. Finally, we show through a running example how these modules can be instantiated.

Paper PDF

Park Vossenberg is a nursing home in the Netherlands that recently changed their care management for residents with dementia from a closed to an open setting. All residents live in studio apartments and are free to move around the nursing home grounds and beyond. In a unique opportunity, a group of researchers was invited to assess the effects of these changes on several involved parties: residents, family members, staff and management, as well as people living in the neighbourhood of the care home. To cover these different effects, data was collected by observations, from wearable and fixed sensors, and questionnaires.

In this progress report, we will focus on the data collection protocols and practical challenges we faced when collecting movement data with wearable sensors and from the nursing home's alarm system. Furthermore, we will present our initial ideas on pattern mining for data exploration and integration.

During medical check-ups, commonly a large variety of measurements, procedures and medication prescriptions is collected. The type of measurements and procedures, however, may differ from patient to patient, hence a dataset can contain a large amount of incomplete samples with respect to the full set of considered features. Common work either use imputation methods (filling in the missing values with some default) or marginalization (ignoring samples with missing entries), which can lead to significant bias or limit generalization. In this work, we present a novel class of adaptive deep neural networks called AnyNets that can deal with missing values by design without the need for imputation. We show the advantage of AnyNets for both supervised and unsupervised learning on various medical datasets.

Paper PDF

This paper describes finding Body Function (BF) mentions within clinical text. Body Function is noted in clinical documents to provide information on potential pathologies within underlying body systems or structures. BF mentions are embedded in highly formatted structures where the formats include implied scoping boundaries that confound existing NLP segmentation and document decomposition techniques. We have created two extraction systems: a dictionary lookup rule-based version, and a conditional random field (CRF) approach based on training from manual annotations. Training and test data utilized the NIH Clinical Center Rehabilitation Medicine Department records. Results of these systems provide a baseline for future work to improve document decomposition techniques.

Paper PDF

Patient-reported outcomes (PRO) are important measures of quality of life for chronically ill patients. These indicators are often difficult to access because they are documented in the clinical notes. We propose a natural language processing pipeline to extract PROs developed using data from a national multi-electronic health record (EHR) registry. Our approach is rule-based and general enough to successfully extract PROs from notes coming from more than 100 practices using more than 20 different EHR brands, demonstrating a great generalizability and potential for transportability.

Paper PDF

The field of Machine Learning (ML) presents tremendous opportunity to develop tools that can optimize decisions involving detection and diagnosis of disease and secondary conditions and address disparities in health and health services. Unfortunately, much of the data generated in the biomedical research are informed by narrowly defined hypotheses and often unamenable to ML techniques owing to poor accessibility, low quality, and incompleteness. Bridge to Artificial Intelligence (Bridge2AI), a National Institutes of Health (NIH) Common Fund program, seeks to address these challenges through a series of grand challenges anticipated to spur widespread adoption of artificial intelligence (AI) and accelerate biomedical advances in health. The program will (1) generate new flagship biomedical and behavioral data sets that are ethically sourced, trustworthy, well-defined, and accessible, (2) develop tools and standards to unify data attributes across multiple data sources and across data types, (3) create automated tools to accelerate the creation of FAIR (Findable, Accessible, Interoperable, and Reusable) and ethically sourced data sets, (4) provide resources to disseminate data, ethical principles, tools, and best practices, and (5) create training materials and activities for workforce development that bridges the AI, biomedical, and behavioral research communities. AI4Function researchers have a great opportunity to formulate grand challenges related to the reliable and valid capture of social, behavioral, and environmental data that associate with well-being and quality of life. Related advances will help transform detection and diagnosis of secondary risks and optimize care, including training and functional recovery for people with disability and chronic morbidities.