|Machine Learning Applications for Improving Behavior Analyst Decision-Making in Practice and Research
|Saturday, May 28, 2022
|12:00 PM–12:50 PM
|Meeting Level 2; Room 252A
|Area: DDA/AUT; Domain: Applied Research
|Chair: John E. Staubitz (Vanderbilt University Medical Center, TRIAD)
|CE Instructor: John E. Staubitz, M.Ed.
Machine learning holds great promise for improving behavior analytic practice and research (Turgeon & Lanovaz, 2020). Historically, behavior analysts have collected and analyzed data as a means of making decisions to improve socially significant client outcomes. When analyzing large data sets reflecting organizations-wide outcomes or the complex outputs that can be captured by sensors, there is a possibility for enhancing the decision-making of behavior analysts. Response effort can limit the extent to which humans can complete analyses or make predictions in time to be beneficial. By nature, machine learning can allow for rapid or even real-time analyses that would be impossible for a human. The first presentation will describe how behavior analysts at an educational center are using sensors to collect physiological and behavioral data and applying machine learning to analyze data and inform decision-making. The second presentation will share data from a multimodal model of sensor data collection and machine learning that allows for real-time prediction of behavioral escalation within a modified Practical Functional Assessment. Finally, presenters will describe a machine learning model for analyzing service delivery and satisfaction data across many organizations over the course of multiple years that allows for improvement in organizational decision-making models.
|Instruction Level: Intermediate
|Keyword(s): Decision-making, Electrodermal Activity, Machine learning, sensors
The target audience for this session includes practicing behavior analysts who oversee behavior change programs that address severe problem behavior, or who oversee ABA agencies and are responsible for making organization-level decisions. This session is also intended to be of interest to those interested in the practical or ethical context surrounding the use of machine learning and or sensors.
|Learning Objectives: At the conclusion of the presentation, participants will be able to: (1) Identify ways in which existing technology can enhance their behavior analytic practice (2) Demonstrate basic understanding of how machine learning and signal processing approaches may be helpful to behavior analysts in the future (3) Demonstrate an understanding of the extent to which using a structured, non-dangerous assessment context may limit assessment time, risk, and resources (4) Describe three strengths and limitations to using machine learning to predict patient outcomes
Integrating Traditional Behavior Analytic Practices With Emerging Technology to Understand and Treat Challenging Behaviors
|JOHANNA F LANTZ (The Center for Discovery), Tania Villavicencio (The Center for Discovery), Corey Olvera (The Center for Discovery), Ali Rad (Emory University School of Medicine, Department of Biomedical Informatics)
Behavior analysts understand behavior through observation of learners in the environment. Technological advances offer a view of what is happening inside of the learner as well. The presenter will describe a program at The Center for Discovery (TCFD) where students in a specially equipped classroom wear sensors that track physiological and movement data. Video data from this naturalistic setting are aligned with sensor data. The presenter will explain how behavior analysts from TCFD integrate data from the sensors with traditional ABA methodology to design better treatments for learners with autism spectrum disorder and maladaptive behaviors. Physiological and behavioral data representing significant clinical findings will be shared. In addition to using technology to inform clinical decisions, the presenter will describe collaborations between TCFD and computer scientists. These scientists are using machine learning and biomedical signal analysis to analyze TCFD’s rich dataset with the desired outcome of automatic detection and prediction of behaviors. The ultimate goal of this relationship is to develop technology that a.) sends alerts to caregivers that a behavior is imminent or that it is time to re-engage following a behavior and b.) detects behaviors automatically as a potential replacement for live data collection.
|Predicting Problem Behavior through a Multimodal Machine Learning-Based Predictive Framework
|JOHN E. STAUBITZ (Vanderbilt University Medical Center, TRIAD), Zhaobo Zheng (Vanderbilt University), Lauren Shibley (VUMC: TRIAD), Nibraas Khan (Vanderbilt University), Amy Weitlauf (Vanderbilt University Medical Center), David Reichley (Vanderbilt University Medical Center), Johanna Staubitz (Vanderbilt University), Nilanjan Sarkar (Vanderbilt University School of Engineering)
|Abstract: Previous research has established the potential for machine learning and physiological data to enhance evidence-based practices for assessing problem behavior. While investigators have demonstrated the capability to predict problem behavior, there are limits to predictive precision, and the assessments needed to build such a model have been time- and resource-intensive, requiring repeated exposures to behavior that poses safety risks to the learner or assessor (Ozdenizci et al., 2018; Goodwin et al., 2019). The practical functional assessment (PFA) allows assessors to efficiently evoke a high number of non-dangerous precursor behaviors in a short period of time. By integrating direct observation data with multimodal data from several sensors capturing the physiological and motion performance of the learner within a modified PFA, we were able to create a model that predicts behavioral escalation with 98.5% accuracy after 1-2 brief assessment sessions. We discuss our process for developing an integrated hardware and software platform with the specific goal of enhancing evidence-based practice in ABA. Additionally, we will connect this promising technology with our existing code of ethics, especially as it relates to minimizing client risk and ensuring informed consent for engagement with technologies that are new.
A Machine Learning Analysis of Applied Behavior Analysis Service Delivery Characteristics That Predict Improved Patient Outcomes
|DAVID J. COX (Behavioral Health Center of Excellence; Endicott College), Zachary Harrison Morford (Texas Association for Behavior Analysis), Jacob Sosine (Behavioral Health Center of Excellence), Cora Gnikobou (Behavioral Health Center of Excellence)
The delivery of ABA services involves a complex interaction of behavioral systems. Patients need to be interested in and seek out ABA services; and, once in ABA, to continue improving their quality of life. Employees need to be hired in sufficient numbers, properly trained, adequately resourced, and appropriately matched with patients they are competent to serve. And payors need to see progress being made within reasonable costs and time frames. In this presentation, we discuss how the Behavioral Health Center of Excellence is leveraging machine learning to describe and understand these complex and interacting behavioral systems. Specifically, we leveraged data collected from 500+ organizations over six years to analyze the interaction between ABA organizations’ systems and processes; staff satisfaction and turnover; service delivery (e.g., utilization rates, hours of ABA contacted); and patient satisfaction with ABA services. We also discuss how these data can predict patient reported progress, improvement in quality of life, and changes on norm-referenced and criterion-referenced assessments. This presentation provides a first look at the variables that might be important to describe and improve the complex interaction of behavioral systems that comprise ABA service delivery.