Machine Learning Tool Enhances Genetic Testing Decisions in NICU, Improving Patient Care Efficiency

A groundbreaking clinical study at Rady Children's Hospital in San Diego (RCHSD) has demonstrated the successful implementation of a machine learning tool to optimize genetic testing decisions in the neonatal intensive care unit (NICU), potentially transforming the approach to diagnosing genetic conditions in newborns.

The prospective study, conducted in the hospital's Level IV NICU, utilized the Mendelian Phenotype Search Engine (MPSE), an innovative tool that employs Human Phenotype Ontology (HPO) terms to assess the likelihood of underlying genetic conditions in patients.

Study Design and Implementation

The research was structured in two distinct phases spanning 52 weeks from July 2022 to July 2023. During Phase 1, which lasted 14 weeks, physicians followed standard protocols for nominating patients for whole genome sequencing (WGS). The second phase, extending 38 weeks, introduced MPSE score reports to aid physician decision-making during daily rounds.

The MPSE tool employs a Naïve Bayes classifier, which has demonstrated robust performance with area under the curve (AUC) values of 0.86 at RCHSD and 0.85 at the University of Utah. The system automatically computed scores every three hours, analyzing clinical notes through natural language processing to generate HPO-based phenotype descriptions.

Significant Enrollment Increases

The implementation showed marked improvements in patient identification and enrollment:

• Phase 1 (14 weeks):

  • 204 eligible infants
  • 27 physician nominations
  • 25 study enrollments

• Phase 2 (38 weeks):

  • 691 eligible infants
  • 91 physician nominations
  • 74 study enrollments

Technical Framework

The MPSE system calculates scores using a training dataset of 1,049 NICU patients, including 293 positive cases and 756 negative cases. The tool generates both raw scores and percentiles, making the results more intuitive for clinical interpretation.

Patient Selection Criteria

The study maintained strict inclusion and exclusion criteria to ensure appropriate patient selection. Eligible patients included NICU admits aged 0-12 months, either within 0-7 days from admission or within one week of developing an abnormal response to standard therapy.

Exclusion criteria encompassed cases with clinical courses explained by:

• Isolated prematurity
• Isolated unconjugated hyperbilirubinemia
• Normal-response infections
• Pre-existing genetic diagnoses
• Isolated transient tachypnea
• Meconium aspiration
• Trauma

Statistical Methodology

The research team employed robust statistical methods, utilizing Python 3.10.2 with specialized libraries. Non-parametric tests were chosen for analysis, including:

• Mann-Whitney-Wilcoxon two-sided tests for group statistics
• Fisher's Exact test for proportion comparisons
• Cox's proportional hazard model for analyzing nomination timing differences between phases

This implementation represents a significant step forward in utilizing artificial intelligence to support clinical decision-making in neonatal care, potentially leading to more timely and appropriate genetic testing for newborns who need it most.