Description
Traditional healthcare is and has been mostly personalized: A patient goes to the doctor after showing some symptoms, the patient is diagnosed with a disease and from then the doctor tries to figure out over time the best treatment as shown in Figure 1. This last stage, can take from weeks to months since access to health professionals is scarce. In this project I argue that we can perform this personalization of treatment in a completely unsupervised manner (without health professional's intervention) by using a combination of AI, sensor data and human feedback.
Figure 1. Diagram showing the process of personalization. At the top the process from the patient's point of view. Middle the health professional's point of view. At the bottom the proposed mobile health solution and its respective challenges as questions.
The main idea as shown in Figure 2. is to use an AI model to determine what and when to deliver treatment using sensor data of health related outcomes and the patient's feedback. As a first step in this project, I explored how to use behavioral data ( e.g., amount of sleep, physical activity, socialization levels, self reported stress) with my own adaptation of Q-Learning (a reinforcement learning method for solving sequential decision making problems). I used the studentLife dataset and used college students self reported and sensed behaviors to determine how their daily activities were affecting their mood and energy levels. Using this approach, I found that too much exercise or too little sleep, exercise or socialization is detrimental for mood and energy levels. I also found that Q-Learning despite my adaptation, has a very high sample complexity (requires too much data) and could not learn fast enough to be used for personalization of a mobile health intervention.
Figure 2. Diagram showing an app that personalize health using the phone onboard sensors a wearable and the patient's feedback.
From this lessons learned, I moved on to my most recent work: a mobile health intervention that adapts the time of delivery of treatment and the treatment itself using an AI method that weights recent behavioral data of the user and the user's feedback to determine the best course of treatment. This approach allows the system to correct itself and learn from experience daily with the patient. This work produced very positive and significant results like improving the main health outcome of our intervention, increasing adherence to treatment and motivation. This work is currently in submission at a top HCI conference. Currently I'm working on novel and fast-to-adapt methods for personalization of timing of treatment. Last, I'm also working on an article discussing how to adapt mobile health interventions for a pandemic.