Background & Aims

For neonates with urgent surgical needs and Intensive Care Unit (ICU) admission, healthcare professionals’ decisions and actions, including crucial painful procedures, significantly impact the newborn’s condition and survival. Pharmacological pain management, notably with opioids, adds a risk of mortality (1). Close monitoring of patient data, stored in Electronic Health Records (EHR), enables professionals to assess responses, predict recovery, and adjust care. However, the intricate attention to these data could be challenging for healthcare professionals due to time constraints, organizational issues, or competency gaps. Machine learning (ML) frameworks could offer efficient solutions for outcome prediction (2). This study aims to identify and assess predictive factors, including pain management strategies, that enhance the likelihood of a child’s survival.

Methods

For this study, the Multidisciplinary Clinical Hospital for Mothers and Children, named after Prof. M.F. Rudnev (Dnipro, Ukraine), provided a dataset of newborns’ medical records, capturing daily descriptions of their medical conditions. The data analysis comprised two phases. In the initial phase, different pre-processing techniques were applied to adapt and transform for ML modeling the manually entered data, including handling missing values, eliminating outliers and redundant features, and encoding categorical data. The preprocessing phase led to a final dataset comprising records of 250 newborns who survived and 198 who did not. The second phase involved developing an ML framework to predict a child’s daily outcome using EHR data. Eight classification models were built employing the XGBoost algorithm, each relative to a different age (0-7 days). The algorithm was chosen within a preliminary modeling study, based on parameter values’ stability across various training sets.

Results

The built prediction models showed an average precision score of 83.37% ? 0.07% (mean ? std), highlighting the discriminative power of EHR data in the first seven days of a newborn’s life. The analysis revealed unique daily metrics influencing outcomes, aligning with the medical community’s emphasis on meticulous day-by-day data evaluation. Notably, chest circumference had the strongest discriminating power on the child’s outcome on day zero in the ICU. Days one and two mirrored each other, introducing the chest and head circumference, the principal diagnoses, and the patient response to examination as influential features. Day three emphasized muscle tone’s significance, while day four highlighted the importance of parenteral nutrition and respiratory patterns. Features like nasogastric tube presence, head circumference, and severity of illness were consistently important for the outcome prediction in at least 50% of the days.

Conclusions

This study highlights the ability to predict newborn outcomes using ML techniques on longitudinal data, emphasizing the importance of standardizing data collection and employing pertinent data pre-processing step for reliable prediction accuracy in medical data-driven decision-making support. The observed limited overlap of only three among the top ten features for each day model reinforces that daily metrics uniquely contribute to outcome determination. Importantly, the identified significant features align with medical knowledge, validating the model’s predictive foundation rooted in hazard assessment. Furthermore, the study underscores that the actions of medical professionals, especially diagnoses, are the most reliable predictors of a child’s survival, offering reassurance that interventions are conducted effectively.

References

(1)Leyenaar JK, Schaefer AP, Wasserman JR, Moen EL, O’Malley AJ, Goodman DC. Infant Mortality Associated With Prenatal Opioid Exposure. JAMA Pediatr. 2021 Jul 1;175(7):706-714. doi: 10.1001/jamapediatrics.2020.6364. PMID: 33843963; PMCID: PMC8042571.

(2)Anita Preininger Fei Wang. “AI in Health: State of the Art”. In: (2019). DOI: 10.1055/s-0039-1677908. URL: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6697503/.

Presenting Author

Natalya Korogod

Poster Authors

Natalya Korogod

PhD

HESAV, HES-SO

Lead Author

Topics

  • Pain in Special Populations: Infants/Children