One of the biggest risks in healthcare is defaulter patients or patients who fail to return for treatment. Timely identification of patients at high risk of defaulting is important and many cases it is collected by hand at the point of care. This data many times cannot be analyzed in real time and there is often a significant time gap of up to a year. This activity uses predictive machine learning algorithms to help front line healthcare workers to identify patients who may default. The work was carried out in India and in South Africa.
This Data Innovation Fund project addressed SDG 3 - Good health and well-being. In this project Dimagi applied machine learning and data analytics to mobile health systems to gain further insight into the dependencies between frontline worker behaviors, mobile data system design, and the process of developing machine learning algorithms and analytics for risk prediction, in this case regarding when patients will default from care. Some key takeaways from this process were:
- Live operation monitoring: This is of high-value when considering the feedback loop between the mobile user and web-based data collection, and using this perspective to develop correlations linking end-user behaviors and priority care delivery metrics.
- Streamlining target development:
- In South Africa, findings showed that patient population subgroups and relevance of appointment types could provide more precise insight into the more general “defaulter” problem in a given population. This has been useful specifically in the treatment of HIV positive patients who need to be tracked for purposes of medication adherence, coinfection monitoring and reduction of transmission.
- Likewise in India, when a delivery place (home vs facility) and infant outcomes (measured by low birth weight) were considered as unique targets, they showed the potential impact of managing different risks within the same population.
- Adding value to on the ground data usage: Understanding how a local NGO engages with data needs and acted on integrating these needs showed the potential to add value and efficiency to high-priority data needs, for improving patient outcomes and strengthening health systems. as well as uncover data “blind spots.” While sophisticated prediction algorithms can be of value, addressing current data needs shows the potential for a high return on effort in terms of its application.
- Machine learning prediction models and data perspectives: Population specific insights were gained using patient-level data and ecological block-level patterns to inform patient outcomes in key areas of maternal health and infant mortality.
Related Links:
Disclaimer: As many of the links are external, some links may not work.