AI-assisted multimodal digital remote screening for bulbar ALS to support timely referral to SLP care

PROJECT SUMMARY The overall aim of this research is to develop novel clinical predictive algorithms (CPAs) to improve the management of ALS via detection of early signs of bulbar disease and prognostication of the risk of clinically significant bulbar indicators. Currently, the onset of bulbar symptoms is not tracked outside of the clinic, which leads to delayed detection of bulbar abnormalities, missing of the critical points in ALS progression, and delayed referral to speech language pathology (SLP) services. There is a critical need to develop a tool for automatic remote assessment to achieve these clinical goals. This research will follow the phased CPA development framework – focusing on its development phase - in service of building robust and valid CPAs for real-world clinical use. In this project, we will address three specific aims (SAs). In SA1, we will examine the effect of repeated test administration (practice) on key features of bulbar dysfunction, such as alterations in speech timing in connected speech. We hypothesize that there will be a practice effect on repeated speech stimuli, but it will level off with time. In SA2, we will develop and internally validate a CPA for screening early bulbar dysfunction. We expect a strong classification performance on the CPA based on multimodal (audio + video) feature set. In SA3, we will develop and internally validate a CPA for predicting the risk of decline on established indicators of bulbar disease progression (speech and swallowing loss) at 6-, 9- and 12-months post-baseline. We anticipate a strong ability of the risk CPA to predict bulbar decline. We will recruit 100 patients with ALS and early stages of bulbar decline and 100 demographically matched healthy controls (HCs). Oro-motor and speech data will be collected using multimodal, web-based platform, which will also be used to generate acoustic and kinematic feature set from audio and video data. The data will be collected longitudinally. We will assess patients and HCs for 5 months, with additional clinical follow-up for patients at 6-, 9-, and 12-months. To address SA1, we will use piecewise mixed- effects models (PMEs) for quantifying practice effects and for identifying potential plateaus. To address SA2, we will use a supervised machine learning (ML) classification model that will be trained to automatically distinguish patients in early stages of ALS from HC. To address SA3, we will use a joint model, which combines a longitudinal model and a survival model, to estimate the probabilistic risk of future clinically significant bulbar change at the three horizons post-baseline. This effort will yield robust, valid, and meaningful CPAs that will be ready for further evaluation (extremal validation) in the context of multisite clinical trials. This work is poised to provide a critical advancement in the clinical management of ALS patients via remote automatic assessment and supply additional valuable information that may be used for surrogate endpoint development for clinical trials.