Ability of an Artificial Neural Network (ANN) to Predict Paediatric Severe Acute Malnutrition (SAM) from Routine Growth Curves

Background: Severe acute malnutrition (SAM) adversely affects many South African children. Growth faltering often precedes SAM, yet is frequently missed during routine growth monitoring (GM). Digitisation of GM may improve detection of growth faltering and prompt earlier intervention to avert SAM, but this requires a locally validated method to identify at-risk children.This study aimed to determine the predictive validity of weight-for-age (WFA) growth curves, assessed by an artificial neural network (ANN, trained with inputs from local child nutrition experts using a multilayer perceptron approach with backpropagation), as predictor of SAM risk in children under five, compared with other indicators of SAM risk. Methods: Among children (under five years old) diagnosed with SAM (n=63) or without SAM (n=122), three approaches were used to classify their routine WFA growth curves as indicators of SAM risk: (1) a locally developed ANN, (2) a logistic regression-derived predictive equation (cross-validated in the study sample), and (3) changes in weight (stagnation/decrease) or z-scores (decrease >0, >0·33, >0·50 and >0·67 z-scores) between the second-to-last and last weight measurements before SAM onset. Diagnostic accuracy testing was performed for all indicators, with SAM risk as exposure and diagnosis of SAM as outcome. Findings: The ANN had sensitivity 73·0% (95%CI: 60·3;83·4%), specificity 86·1% (95%CI: 78·6;91·7%) and receiver operating characteristic (ROC) area 0·795 (95%CI: 0·732;0·859). The predictive equation had similar sensitivity (73·0%, 95%CI: 60·3;83·4%), but lower specificity (50·8%, 95%CI: 41·6;60·0%) and ROC area (0·619, 95%CI: 0·548;0·690). Indicators based on changes in weight or WFA z-scores produced ROC areas between 0·634 and 0·677. Interpretation: The ANN had the best predictive validity of the indicators considered, but would benefit from further refinement. Careful consideration of costs and benefits to the local healthcare system should precede its inclusion in digital GM. Funding Statement: No external funding was received. Declaration of Interests: None of the authors have any conflict of interest to declare. Ethics Approval Statement: Ethical approval (Faculty of Health Sciences Research Ethics Committee, University of Pretoria: 468-2017), and permission to collect data were obtained from health authorities and the health institutions involved. Informed consent was obtained from the parent/caregiver of all participants.