Development of a prediction model for infant hospitalisation and death using clinical features assessed by community health workers during routine postnatal home visits in Dhaka, Bangladesh
Alastair Fung, Marimuthu Sappani, Cole Heasley, Chun-Yuan Chen, Shaun K Morris, Peter J Gill, Diego G Bassani, Davidson H Hamer, Prakesh S Shah, S M Abdul Gaffar, Sultana Yeasmin, Shafiqul A Sarker, Shamima Sultana, Joseph Beyene, Daniel E RothBackground
To improve upon the WHO 8 danger signs used to identify young infants (<2 months) requiring referral during community health worker (CHW) home visits, aggregative features (eg, cumulative visits with fever) rather than visit-specific features (eg, fever at a single visit), and a machine learning random forest model, may enhance predictive performance. Applying these approaches, we aimed to develop a prediction model for infant hospitalisation and/or death using CHW-assessed clinical features during home visits in Dhaka, Bangladesh.
Methods
We analysed data from generally healthy infants prospectively enrolled at birth and assessed at 11 scheduled CHW visits from 3 to 60 days of age. To predict first hospitalisation or death, we developed two models—time-varying Cox regression and random forest—using the same candidate predictors (45 clinical features of which eight were WHO danger signs and 12 additional covariates) with aggregative features incorporated. We evaluated discrimination (C-statistic) and calibration (calibration plots). Performance was compared with a time-varying Cox model using only WHO danger signs.
Results
Among 1906 infants, 176 (9.2%) had an event (173 hospitalisations, three deaths). The best-performing Cox model (C-statistic=0.71; 95% CI 0.68 to 0.75) consisting of three baseline covariates (any perinatal/delivery complication, umbilical cord care and gestational age) and four visit-specific clinical features (nasal congestion, cough, jaundice and skin rash) and a Cox model with these four features plus WHO danger signs (C-statistic=0.70; 95% CI 0.67 to 0.74) demonstrated higher discrimination than WHO danger signs alone (C-statistic=0.56; 95% CI 0.54 to 0.60), with similar calibration. A random forest model (42 predictors) was well-calibrated with comparable discrimination (C-statistic=0.69; 95% CI 0.64 to 0.73).
Conclusions
Aggregative features and random forest did not outperform a time-varying Cox model using baseline covariates and visit-specific features. Among lower-risk infants, adding four features to WHO danger signs may improve predictive performance by capturing a broader spectrum of illnesses requiring hospitalisation.