Project 458245

In low- and middle-income countries (LMICs), current World Health Organization (WHO) guidelines recommend routine home visits by community health workers to assess infants and urgently refer to hospital if any of eight 'danger' signs (e.g., fever >37.5C) are present. However, these 'danger' signs came from studies in which infants presented to hospital due to caregiver concern. A prediction model developed among infants assessed during home visits, rather than at the hospital, can utilize repeated measurements of clinical features over many home visits. Using repeated measurements may allow for earlier identification of illness. It also allows for an infant's temperature, for example, to be compared to his/her own norm based on prior measurements, which may predict illness more accurately than using a general threshold like 'fever >37.5C.' Highly accurate machine learning models have not been used to predict illness in infants in LMICs. Our goal is to investigate how accurately clinical features assessed during home visits can predict infant hospitalization and death in Dhaka, Bangladesh. We will develop two prediction models, one using a conventional method called regression and a second using a machine learning approach. The models will be developed using data from a clinical trial in Dhaka. We will compare the accuracy of the two models and verify whether they perform accurately in a separate group of infants from a different clinical trial in Dhaka. Per WHO guidelines, any infant 'danger' sign detected on a single home visit signals the need for hospital referral which may lead to over-referral and overburden low-resource health systems. Using novel and accurate machine learning methods and establishing patterns of concerning clinical features detected over many home visits can identify at-risk infants early, prevent severe illness and prevent unnecessary referrals. Our findings may contribute to improving WHO guidelines and reducing infant mortality in LMICs.