Deep Detection: Statistical Modeling and Deep Learning to Predict and Address Malaria and Anemia in Madagascar
MetadataShow full item record
CitationAnnapragada, Akshaya. 2020. Deep Detection: Statistical Modeling and Deep Learning to Predict and Address Malaria and Anemia in Madagascar. Bachelor's thesis, Harvard College.
AbstractIn Madagascar, malaria and anemia are significant public health challenges. Despite global control efforts, malaria prevalence in Madagascar remains high (estimated at 7% nationally, but approaching 30% regionally) and has increased in recent years. Anemia prevalence is similarly high (half of children under five and over 30% of women aged 15-49). There are well-established links between malaria and anemia, with untreated malaria being a cause of anemia, and the co-incidence of the two diseases driving substantial morbidity and mortality. These disease burdens exist against a background of high poverty and malnutrition, limited access to health care, and weak disease monitoring infrastructure.
I analyzed a survey dataset documenting socio-economic, environmental, and dietary variables, and malaria/anemia disease status for 5405 individuals across 24 sites in Madagascar, and built models to predict malaria, anemia and co-incidence disease status. I then applied interpretability methods to identify variables important to disease status classification and analyzed these variables from a public health perspective.
I found that neural network models predicted malaria, anemia and co-incidence more effectively than logistic regression or principal component analysis-logistic regression models, though logistic regression models are more easily interpretable. The best performing neural network models had >70% recall, and >80%, >70%, and >85% accuracy for malaria, anemia and co-incidence respectively. These models could enable patients far from health clinics to be screened for malaria, anemia, and co-incidence using only easily assessed, non-clinical variables, hence enabling more efficient interventions in the context of limited health care availability.
Moreover, I found that most variables identified as important to disease status classification were socio-economic factors, including water and sanitation practices, dietary variables with connections to poverty, and access to health care services. These results suggest that for maximal impact, malaria and anemia interventions will need to be broad, multi-sectoral programs that address poverty and its accompanying circumstances.
Citable link to this pagehttps://nrs.harvard.edu/URN-3:HUL.INSTREPOS:37364672
- FAS Theses and Dissertations