Clinical trials across the US and the UK demonstrate first early detection of Alzheimer’s disease with AI-based speech analysis
Released October 17, 2022.
Results from two clinical trials published today in the scientific journal Brain Communications provide evidence that very early stages of Alzheimer’s disease can be detected through the way people speak using AI-based speech analysis. The speech-based system could detect the presence or absence of brain amyloid protein at the very early stages of Alzheimer’s disease (preclinical and prodromal stage), typically requiring PET scans, and did so better than the Preclinical Alzheimer’s Cognitive Composite with semantic processing (the “PACC5”), a sensitive clinical measure of subtle cognitive impairment that requires long in-person administration by highly trained staff.
Importantly the speech-based system was fully automated and can be run in a browser on any standard mobile, tablet, or computer. It could therefore offer a more accessible and affordable method for testing for early Alzheimer’s at scale.
One of the underlying disease processes of Alzheimer’s disease, accumulation of amyloid protein in the brain, begins up to decades before usual diagnosis. Amyloid protein has been a key target for new Alzheimer’s treatments, including the drug Lecanemab that was recently the first disease-modifying treatment to demonstrate positive treatment effect on cognitive decline. Clinical trials typically use PET scans to detect amyloid protein. But with a cost of thousands of dollars per PET scan, this method has been limited to clinical research. In the broader healthcare system, clinicians most often rely on simple pen-and-paper-based tests with limited ability to detect early cognitive decline, causing long delays in diagnosis.
Now researchers from Novoic, Harvard Medical School and Strategic Global R&D have demonstrated that the presence of amyloid protein in the brain can be detected during the earliest stages of Alzheimer’s disease from how people speak, using AI-based speech analysis, potentially reducing the time to diagnosis and the costs to detect the presence of amyloid protein.
The published data comes from two recently completed clinical trials in the UK and the US run by Novoic: Amyloid Prediction in Early Stage Alzheimer’s Disease From Acoustic and Linguistic Patterns of Speech (NCT04828122, NCT04928976). Whereas previous work has considered speech changes in dementia, the AMYPRED trials focussed specifically on early stage cognitive impairment (preclinical and prodromal stage), where the presence or absence of amyloid biomarkers had been confirmed as the underlying cause – matching the disease-stages where new disease-modifying treatments such as Aducanumab, Lecanemab and Donanemab are being tested today.
“We set up the AMYPRED trials to test if there is a speech phenotype at the earliest stages of Alzheimer’s disease, and whether there are speech differences in amyloid biomarker positive individuals, who may be at greater risk for developing cognitive problems or Alzheimer’s Dementia in the future” said Emil Fristed, CEO of Novoic, “Using a novel class of clinically informed deep learning models we reveal a distinct speech phenotype at both the prodromal and preclinical stage.”
To test patients’ speech, the researchers used a recently published speech-testing protocol called the Automated Story Recall Task (ASRT) (Skirrow 2022, JMIR Ageing), where patients are played one or more pre-recorded short stories via a smartphone or computer and asked to tell them back. The spoken responses are recorded and can then be analysed for episodic memory function and language use.
Turning to the task of predicting the presence of amyloid biomarkers in the brain from the spoken responses, the researchers found that the methods traditionally used for speech analysis had limited use. Instead, they developed an approach based on a recent breakthrough in natural language processing, in the area of large language models, and applied the current state-of-the-art deep learning model for evaluating text pairs (Weston 2022, ACL) to the spoken responses. Novoic’s CTO, Dr. Jack Weston, explained:
“Traditional statistics extracted from speech lack the sensitivity we need to detect changes associated with the earliest stages of the disease. Off-the-shelf foundation models, like BERT and its derivatives, also don’t cut it when we’re dealing with dataset sizes typical of clinical research. Instead, we developed a custom AI model primed to detect clinically-relevant patterns, fusing the latest research in clinical neuroscience and large language models. When we used this model to evaluate the quality, content and form of the responses to predict the participant’s amyloid biomarker status, the model performed remarkably well.”
Using this model, the researchers found that they were able to detect the presence or absence of brain amyloid protein at the very early stages of Alzheimer’s disease. The speech-based system outperformed the Preclinical Alzheimer’s Cognitive Composite with semantic processing (the “PACC5”), a sensitive clinical measure of subtle cognitive impairment that requires long in-person administration by highly trained staff, and which is used as an endpoint in several clinical trials today in preclinical Alzheimer’s disease.
Importantly the speech-based system was fully automated and can be run in a browser on any standard mobile, tablet, or computer. It could therefore offer a more accessible and affordable method for testing for early Alzheimer’s at scale.
The system is already being rolled out across the US, including as the digital screener for the fourth generation of the Alzheimer’s Disease Neuroimaging Initiative (ADNI4) – the largest and most diverse recruitment effort in early stage Alzheimer’s disease to date. Here Novoic’s Storyteller system will support screening of 20,000 patients online to inform participant selection across >50 sites, providing an early example of the scalability of the system.
The Automated Story Recall Task (ASRT) and the accompanying automated AI-based analysis, marketed together as “Storyteller by Novoic”, is available for researchers and for commercial use from Novoic.
About Novoic
Novoic is an Oxford-founded clinical stage digital medtech company developing clinical deep learning algorithms to detect early-stage neurological disease from how people speak. The company’s AMYPRED clinical studies were the first to test deep speech phenotyping in biomarker-confirmed early-stage Alzheimer’s disease, and its research team has developed the state-of-the-art clinical deep learning model for text evaluation. Novoic is working with organisations such as the Alzheimer’s Disease Neuroimaging Initiative (ADNI), Alzheimer’s Drug Discovery Foundation (ADDF) and the UK’s National Health Service (NHS) to accelerate clinical translation of speech biomarkers into impactful medical devices.
About Storyteller by Novoic
Storyteller (by Novoic) is an automated system for speech-based cognitive testing that uses AI-based natural language processing to quantify patterns of speech associated with subtle patterns of disease-specific, early cognitive decline. It runs AI-based speech-analysis on Novoic’s Automated Story Recall Task (ASRT), a digital and automatically administered story recall task, that’s been demonstrated to have excellent psychometric properties and good usability for remote self-testing on people’s own devices, including in an elderly, cognitively impaired population (Skirrow 2022, JMIR Ageing). The system runs as a browser-based app on any mobile device or computer and takes less than 10 minutes to complete, and can be used by patients themselves at home on their own devices. Storyteller uses automated quality control, audio processing and transcription, and AI-based analysis, to provide instantaneous results on subtle cognitive decline and risk of amyloid positivity. It’s been validated in a number of gold-standard clinical studies, including the AMYPRED studies (NCT04828122, NCT04928976). The system is already being rolled out across the US in a number of large-scale research efforts, including as the digital screener for the fourth generation of the Alzheimer’s Disease Neuroimaging Initiative (ADNI4) – the largest and most diverse recruitment effort in early stage Alzheimer’s disease to date. Here Novoic’s Storyteller system will support screening of 20,000 patients online to inform participant selection across >50 sites.
About the AMYPRED studies
The AMYPRED studies (NCT04828122, NCT04928976) are the first clinical studies to investigate deep speech phenotyping in a biomarker-confirmed, early stage Alzheimer’s disease population. The objectives of the studies were to test if a novel first-in-class clinical deep learning model can detect speech patterns specific to mild cognitive impairment and to the presence of amyloid biomarkers in the brain. In both the UK and US-based study, participants with mild cognitive impairment and normal cognition, and with confirmed positive and negative amyloid biomarker status, were recruited into four distinct cohorts. In the studies participants did a battery of speech tasks both under supervision of a clinician and remotely with unsupervised self-testing.