457 - A Machine Learning-Based Model for Prediction of Culture-Proven Early-Onset Neonatal Sepsis in Low- and Middle-Income Countries (On behalf of Investigators of the DeNIS collaboration*)

Monday, May 1, 2023

9:30 AM – 11:30 AM ET

Poster Number: 457
Publication Number: 457.432

Deepika Anand, All India Institute of Medical Sciences, New Delhi, India, Rohtak, Haryana, India; Ayushi Aggarwal, Indraprastha Institute of Information Technology, Delhi, Delhi, Delhi, India; Pradeep Singh, Indraprastha Institute of Information Technology, Delhi, delhi, Delhi, India; Mudit Kapoor, Indian Statistical Institute, New Delhi, Delhi, India; Tavpritesh Sethi, Indraprastha Institute of Information Technology Delhi, New Delhi, Delhi, India; Jeeva Sankar, AIIMS, New Delhi, New Delhi, Delhi, India

Presenting Author(s)

Deepika Anand, MD, DM (she/her/hers)

Neonatal Fellow
All India Institute of Medical Sciences, New Delhi, India
Rohtak, Haryana, India

Background: Existing prediction models (machine learning [ML] or otherwise) for timely diagnosis of early-onset sepsis (EOS) are mostly developed from high-income countries. Differences in pathogen profile and antimicrobial resistance patterns preclude their use in low- and middle-income countries (LMICs).

Objective: To develop and validate an ML-model to predict culture-proven EOS in neonates with suspected sepsis.

Design/Methods: We used ‘Delhi Neonatal Infection Study’ (2011-2014) data from three level-3 hospitals. Neonates who underwent evaluation for EOS were included. Maternal/neonatal variables–except those clearly suggesting sepsis, e.g., foul-smelling liquor–were included. An 80:20 data split was used with culture-proven sepsis as reference standard. Feature selection was performed with Boruta wrapper technique, that iteratively eliminates low-importance variables. Five ML models (RF, Gradient boosting, Extreme Gradient boosting, Support Vector Machine, and Adaboost) were developed using Boruta features. Based on area under receiver operating characteristics curve (AUROC), sensitivity, specificity, positive and negative predictive values (PPV and NPV), best model was chosen and incorporated into a web-based application. The target sensitivity was 90%. We externally validated the model using an ambispective cohort (January 2015-September 2021) at AIIMS, Delhi.

Results: Of 3897 neonates evaluated for EOS, 548 had culture-proven sepsis while 2376 had no sepsis (remaining 973 neonates with culture-negative sepsis were excluded). The mean birth weight and gestation were 2112 g and 35 weeks, respectively. The median (IQR) age at suspicion was 13 (5-33) hours. Boruta selection identified a set of 30 high-performing features (Table 1). Among 5 ML models, RF algorithm performed best on test set with AUROC 0.83, 90% sensitivity, 56% specificity, 45% PPV, and 93% NPV and incorporated into a computer application (Figure 1). In external validation, 26 (of 271) neonates had culture-positive and 121 had no sepsis (124 with culture-negative sepsis were excluded). The model had 92% sensitivity and 34% specificity. Assuming all neonates with suspicion were treated, our model would have avoided antibiotics in 41 (27.9%) while delaying treatment in 2 (1.4%) neonates.

Conclusion(s): Arguably, the first ML-based model from LMICs involving large prospectively collected data has good accuracy for predicting culture-proven EOS in neonates.

*H Chellani, R Gaind (VMMC); S Ramji, S Kumar (MAMC); M Jajoo, V Manchanda (CNBC); V Sreenivas, A Kapil, A Thukral, R Agarwal, A Deorari, V Paul (AIIMS); New Delhi