For doctors to diagnose and treat Parkinson’s, they need reliable tests. Unfortunately, testing the symptoms of Parkinson’s at a specialist’s office/clinic is expensive and time-consuming. We have found that voice recordings collected in the clinic contain enough information to detect Parkinson’s, but we want to test the same capability over the phone, or using smartphones, such as an iPhone. We will also test whether it is possible to detect Parkinson’s in the very early stages from voice recordings.
The project is mainly based around data analysis. We will analyze a massive database of pre-recorded phone calls donated by the public, and voice recordings collected in the clinic from people with a certain genetic mutation implicated in Parkinson’s. We will also analyze other data such as body motion, finger tapping and reaction times collected using smartphones from Parkinson’s patients. We will test, for the first time, sophisticated computer programs to make predictions about whether someone has Parkinson’s or not, and if so, at what precise level of symptom severity.
Relevance to Diagnosis/Treatment of Parkinson’s Disease:
Currently, there are no simple, cheap and accurate tools for remote testing of Parkinson’s symptoms. The tools we are developing in this project could rapidly track how symptoms change over time, and, with this information, we can develop personalized treatment regimens based how each individual responds to each treatment. We also envision nationwide screening for Parkinson’s, tracking changing demographics over time.
For voice recordings, because we have so much data, we expect to have the definitive answer to how accurately it is possible to detect Parkinson’s over the phone. We also expect to learn something about whether voice changes are an early sign of the condition. Finally, we will know by the end of the project how feasible it is to detect, and how accurately we can track over time, Parkinson’s symptoms using smartphones such as Androids or iPhones.
In a previous study, we collected behavioral data recorded from Parkinson’s patients and healthy people using a smartphone app. In a project called the ‘Parkinson’s Voice Initiative’, we also collected a very large number of voice recordings (approximately 17,000) over the telephone. In this study funded by the Michael J Fox Foundation, this data was analyzed and processed using machine learning techniques. The aim was to make predictions about the health status of each participant from the behavioral data.
We were able to show very high accuracy in separating healthy people from those with Parkinson’s from the smartphone data (nearly 97% accuracy). We could also predict symptom severity on the UPDRS scale with an error of less than 1 UPDRS point. With the telephone voice data, we achieved accuracy of around 64% in separating healthy people from those with Parkinson’s. This demonstrates that both smartphones and the telephone are useful methods for detecting and quantifying the symptoms of Parkinson’s. Smartphone accuracy was extremely good. Telephone accuracy was, however, not very good. This discrepancy in accuracy reflects the difficulties in collecting voice recordings of sufficiently high quality over the telephone, and further work is underway to develop algorithms which can identify when a voice recording captured over the telephone is useful.