Automated speech assessment for diagnosis of FTD spectrum disorders - PROJECT SUMMARY/ABSTRACT Alzheimer’s Disease and Related Dementias (ADRDs) are major causes of death and disability in the United States, with the number of adults with ADRDs projected to reach 13 million by 2050. Among these, the frontotemporal degeneration (FTD) spectrum disorders are among the most common causes of dementia in younger individuals (<60y), resulting in high social and economic burden. Diagnosis of FTD is challenging, typically requiring subspeciality expertise that is not widely available in a timely manner. Many FTD spectrum disorders manifest in speech, and speech changes can help differentiate FTD subtypes. Prior research supports the clinical utility of speech-based prediction of FTD presence and subtype. Unfortunately, rigor of these prior studies is limited for several reasons, including small sample sizes, failure to follow predictive modeling best practices, use of research grade speech recordings, and lack of prospective validation. We propose a highly innovative approach to speech-based prediction of FTD that avoids the weaknesses of prior work. Central to our approach is the insight that FTD may be too rare to use powerful deep learning models, but the abnormal speech characteristics seen in FTD are also seen in other disorders. Training a model to recognize these characteristics in FTD does not require limiting the dataset to FTD patients. We plan use this to our advantage. In Aim 1 we will use a self-administered, web-based speech exam to create a large dataset of all disorders seen in our speech clinic. Recordings will be made in a standard exam room using mobile phones or tablets. Our expert speech and language pathologist will annotate the recordings with perceptual speech characteristics, such as abnormal rate or vocal strain. The large sample size will enable us to use deep learning for what it excels at – trainable feature extraction optimized for the task at hand. We will follow predictive modeling best practices, including use of a validation set. In Aim 2 we will apply these trained models in a large cohort spanning the FTD spectrum and extract the data from the last layer in the network, just before it makes its prediction. This is a low dimensional representation of the speech signal, but which contains the information necessary for predicting perceptual characteristics. We will use these representations to develop a nearest neighbor classifier for FTD. Essentially, the model matches a new case to similar ones in a labeled set based on the low-dimensional representation and uses the neighborhood to assign a label for the new case. Finally, in Aim 3 we will combine our self-administered speech exam and the models from Aims 1 and 2 into a single tool and perform a prospective validation study to test performance in a clinical setting. The predicted increase in ADRDs and lack of access to specialists will necessitate a shift in clinical practice from a few expert centers to a distributed system of non-expert providers. The digital tool we propose meets this challenge head on through scalable and easy to use automated speech analysis and prediction of FTD.
