As we discussed last week, we’ve narrowed down our area of interest to using language as a method for detecting signs of Primary Progressive Aphasia (PPA) before it becomes problematic. After speaking to Dr. Murray Grossman, a researcher at Penn Medicine, we learned that PPA provides early clues of its onset in the form of subtle changes to the acoustic profiles and word choice of an individual’s speech. Based on this information, we decided to focus on building an app that would allow individuals to detect their level of risk for PPA by collecting voice data. Since making this decision, we’ve started to put some more thought into how our app would work, as well as what types of challenges would be involved in bringing it to fruition.
Overall, our app (called Synaptax) would work by prompting its users to provide a brief summary of their day every week. After collecting the voice data associated with this response, the app would compute an overall risk score and show the user their trend in this risk score over time. Based on the risk score and its trends, the app would also provide each user with automated recommendations regarding whether they should seek medical attention, using location data to list medical experts nearby who would be most equipped to provide relevant medical assistance.
In order to estimate risk, the app would incorporate two separate sets of indicators: those related to how the user’s speech patterns compare to their past speech patterns, and those related to how the user’s speech patterns compare to both the speech of healthy individuals and the speech of individuals at risk for PPA. We’ve chosen to include both types of indicators in the underlying model for our app because we believe they both provide crucial information; only analyzing trends in a user’s speech may be difficult to draw conclusions from without having other data to provide context, but only having data from other patients might make it difficult to discern which aspects of a patient’s speech stem from PPA and which come from their natural speech patterns. To produce a risk score from this data, we would train a machine learning model on various individuals’ histories of speech data and information regarding whether they developed PPA.
Now that we’ve told you about our plans for our app, here are some diagrams of how it would work:
The diagram on the left shows what a potential main page for the app could look like. It gives users the option to record their current speech ("Test"), the option to learn about PPA, and the option to view their "Scores," including their current risk level and their trends in risk level. The middle diagram shows the "Scores" page, which also provides a link to the page on the right, which displays recommendations for the user.