In the last 30 years, advances in computing have allowed even low-resource settings to use innovative analyses of global health measurement. However, there are gaps in the efficacy of existing analytical techniques, including the ability to fill in missing data that span a lengthy time period, appropriately represent uncertainty in results generated from complex methods, and apply a wide range of statistical techniques in ways that are efficient, accurate, and consistent.
The Computational Algorithms research team develops and improves statistical methods, modeling techniques, and programs to help solve complicated analytic problems present in different types of data from national and international health information systems and from locally available sources to advance the science of health measurement and evaluation.
In order to support IHME’s mission to examine the current state of population health and strategies to improve it, the team identifies common methods useful to other IHME research teams, vigorously validates those methods, and provides tools and best practices for their use.
Key Activities
- Create a single user-friendly software package that facilitates spatial-temporal data analysis and handles small numbers and a high degree of sampling variance
We are working to improve upon internal software versions. Since many researchers commonly deal with some of these challenges, the aim is to create software that could be used by a wider body of researchers than just those at IHME.
- Standardize an approach to the Random Forest Method for cause of death classification
The Computational Algorithms research team leverages the vast body of machine learning research produced by computer scientists and statisticians to adapt state-of-the-art techniques to the specific challenges of global health metrics.
One example of how machine learning has impacted IHME’s research is in the area of verbal autopsy. Verbal autopsies are interviews conducted with family members of someone who has died to learn about symptoms present before death. They provide especially useful data in countries with limited or nonexistent vital registration systems. Questionnaire responses form binary, categorical, and continuous attributes, and the disease classification is the categorical response. Disease classification categories typically number above 30 and may be as many as 150. In verbal autopsy, where there are many different causes of death, we can substantially improve traditional approaches to machine learning by taking the hierarchical structure of the causes into account.
The Random Forest Method is one machine learning approach, and it has yielded good results with the verbal autopsy analyses. However, it can be applied in different ways. Therefore, we are working to draft guidelines that will facilitate the application of this technique to other research questions at IHME.
- Standardize an Institute-wide approach to uncertainty, wherein the quantity reported is the same but the methods may vary substantially
A draft set of general guidelines for predictive validity as a tool to validate complex models for global health estimation has been written and is being circulated internally for comment. It will be refined and submitted to a peer-reviewed publication in 2012. Additionally, the research team will develop a tool that generates cross-validation tests to produce predictive validities of algorithms.
Impact
The novel models and methods being developed by the Computational Algorithms research team have the potential to change the way global health research is conducted. The guidelines developed also help IHME to maximize the computational power available in analyses, while ensuring systematic approaches and comparability of results as much as possible.
Progress on the timely, valid, reliable, and comparable measurement of health indicators requires tools and instrument innovation. By using the methods and tools developed by the Computational Algorithms research team, decision-makers in all countries will benefit from more timely, valid, reliable, and comparable measurements of the health constructs that they seek to influence.
