Mobilizing Health (MH) aims to connect rural patients with doctors even though they may live in incredibly remote locations. Through a network of community health workers and cellphone technologies, Mobilizing Health connects millions without healthcare to patient services they otherwise wouldn't have.
Mobilizing Health has rich data on the many cellphone requests made to doctors from community health workers. From this data MH had questions about doctor performance, latency in patient requests, and general symptom trends. We worked with them during our San Francisco Datadive to help answer those questions together.
The team analyzed the response rates for doctors based on when the request was made and when a response was given. Overall, the team didn't find significant differences in response time given time of day or week, which was a superb finding. We could expect certain times of day or week to be worse times to make requests, but overall our team found that the doctor's response time was fairly constant.
Our DWB + Mobilizing Health team examined the types of prescriptions doctors were prescribing over time. The team had to use text processing skills to extract likely prescription names from the unstructured text. This analysis allowed them to analyze when certain drugs were being more heavily prescribed or less heavily prescribed than usual. Such differences can be indicators of epidemics, changes in doctor behavior, or changes in symptoms that patients are reporting. Regardless of the cause, this analysis allows Mobilizing Health to quickly understand when trends are changing and can look more deeply into them.
The figure above shows the yearly and monthly percentage of prescription types prescribed. The third column shows the change that month, which provides an indicator for sudden changes in prescription behavior.
Our team built a visualization of the number of requests going to a doctor versus those coming from that doctor over time. This graph allows MH to quickly identify very responsive doctors or very unresponsive doctors as well as to understand what proportion of each exists in their program. Having this information would allow MH to potentially reroute messages to certain responsive doctors at certain times or if the case is particularly urgent. The figure below shows one month of activity to and from doctors, where red dots indicate messages sent to doctors and blue dots indicate messages returned from doctors.
The DWB + MH team had the amazing realization that, with enough data, they might be able to predict what drug should be prescribed without even contacting the doctor. This finding would be hugely helpful as it would allow MH's systems to learn from previous doctor responses and suggest solutions to requests without even having to consult a doctor. Of course a human would always be involved in the loop for verification, but being able to get a faster recommendation without a doctor's intervention would be hugely beneficial.
To tackle this problem our team started doing analysis of the messages to determine what prescriptions appeared in responses to certain symptoms. The initial findings were very positive and allowed the team to tie certain known symptoms to two prescribed drugs. There is more than one use for this tool and, in the words of the team's Data Ambassador Chris Diehl,
Leveraging the work done to identify trending drugs being prescribed, one can use a trained model for a given drug to identify significant predictive terms and display them. This gives any user of the web application insight into the factors driving the doctor's decisions. This can be especially useful in cases where a given drug has broad-based applicability. The model could be trained online to respond to changing conditions on the ground. A longer-term goal that could be addressed with this same approach is to simplify the doctor's workflow in response. If the predictive models could successfully predict treatment for a large class of common ailments, then the prediction could be made and sent to the doctor for verification. If the predicted treatment is correct, the doctor could simply affirm the proposed treatment. In the worse case, they revert to the standard procedure for specifying a response.
The DWB + MH team was able to answer a range of questions about Mobilizing Health's data that will greatly help them understand how their program is being used. They were able to understand doctor response rates over time, verifying that there were not particularly favorable times of day or week for requests and identifying which doctors were least and most responsive. They also took on the messy problem of dealing with unstructured text and, in the process, created tools for analyzing trends in prescriptions over time as well as predicting prescriptions from given message text. These tools and analyses can be used by Mobilizing Health right now, but are also being continually developed by Chris and his team.