Crowdsourcing health research, analyzing health data, and building personalized AI-based health coaches.
Our 2023 ResearchGate paper showed correlations between exercise and knee pain in three patients with chronic knee pain. Our analysis was based on big data (at the patient level), recorded daily for several years, and was thus the first study to examine those correlations in such fine granular detail.
We showed that exercise increases always preceded pain increases, but, as illustrated in the figure above, the delay between exercise peaks and pain peaks could vary. The long delays might be explained by periods of exercise “build-up” that only impact pain levels later on. The long delays also illustrate the difficulty for chronic pain patients in “choosing” the amount of exercise they should do, as the “pain warnings” to slow down exercising often occur when it's already too late. Therefore, creating an AI to warn patients of incoming danger should be useful.
More recent analyses have also shown that our data can reveal how different types of activities may have different impacts on symptoms, possibly suggesting activities that patients should avoid/increase. We hope to publish those new findings soon.
Here’s a graph generated from our first attempt at creating an AI based on deep learning to deliver warning signs to patients. The final system would enable patients to input their planned activities for the day and see whether those would lead to a warning sign, thus enabling them to make better decisions.
From this graph, we can see that there are still warnings the AI should not have placed, some warnings that are missing, and some ambiguous warnings (not clear whether they should or should not be there). The unpredictability of the delay between exercise bouts and pain bouts, as observed in the previously mentioned 2023 ResearchGate paper, may also be limiting what can be achieved with the mostly behavioral, rather than biological, data currently used. That said, further improvements to this AI system are likely possible with the current behavioral data, and even greater improvements should be possible if biological data were also integrated.
Feel free to contact us at olivier.mirat.om@gmail.com