Journal of Krishna Institute of Medical Sciences University (Oct 2022)

A novel approach to predict the risk of invasive candidiasis using artificial neural networks and comparison with other models

  • Sheetal Gouda,
  • R. Sathyajith,
  • Basavaraj V. Peerapur

Journal volume & issue
Vol. 11, no. 4
pp. 10 – 19

Abstract

Read online

Background: Invasive Candidiasis (IC) is the third common cause of neonatal sepsis associated with high mortality. Low Birth Weight (LBW), prematurity, diabetes, and multifocal colonization increase the risk of IC. Since the development of IC involves multidimensional risk factors, various models are available to predict IC, including colonization index, corrected colonization index, Ostrosky's clinical prediction rule, and candida score. Aim and Objectives: To develop a novel model using Artificial Neural Networks (ANN) that has a strong capability to handle complex multi-variate risk factors. Material and Methods: A prospective, hospital-based observational study was conducted at Raichur, India, among neonates of LBW (<1500 g). Swabs were collected from various body sites followed by blood culture from each neonate to assess colonization and IC, respectively. χ -tests of significance were applied to ascertain significant risk factors for IC. Various risk prediction models were compared with the ANN model. Results: The study population consisted of 103 neonates, of which 21 were diagnosed with IC. The most common isolate obtained was C. albicans, followed by C. parapsilosis. The factors significantly affecting IC were gestation age, mode of delivery, respiratory distress, diabetes, use of antibiotics, and multifocal colonization. ANN model predicted IC with a PPV of 83.3% and a Negative Predictive Value (NPV) of 98.7%. Various prediction models had poor values of PPV and sensitivity compared to the ANN model. Conclusion: ANN showed higher PPV, NPV, sensitivity, positive likelihood ratio with a very low value of negative likelihood ratio, suggesting that ANN is superior to the other models. With multi-centred data available from various geographic regions, the ANN approach can be further strengthened for usage in public and private health setups.

Keywords