Deep learning-based model development to predict critical pediatric events in general wards

In a recent study published in Scientific Reports, researchers used deep learning to develop a model that predicts critical events in pediatric individuals admitted to the general ward using simple variables.

Study: Development of a deep learning model that predicts critical events of pediatric patients admitted to general wards. Image Credit: PopTika/Shutterstock.com

Background

Early diagnosis of deteriorating individuals is vital for prompt management before critical events like cardiopulmonary resuscitation (CPR). Children are more likely to receive treatment before cardiac arrest. Existing tools are time-consuming and complex, rendering them unworkable.

The Pediatric Early Warning Score (PEWS) has poor predictive ability. Deep learning is used to develop prediction models for medical crises; however, most research has focused on adults.

One study employed 29 criteria to determine the ICU transmission probability, which may be unrealistic. Another created an LSTM model that needed >20 vital sign measurements.

About the study

In the current retrospective cross-sectional observational study, researchers developed a machine learning model to predict crucial events in pediatric patients admitted to general wards based on characteristics such as vital signs, age, gender, and measurement interval.

The team conducted the study from January 2020 to December 2022, including patients aged <18 years hospitalized admitted to a tertiary pediatric hospital’s general ward.

They characterized critical events as CPR in general wards, an unanticipated transfer to intensive care units (ICU), or death.

They trained a critical event prediction model using vital signs collected during hospitalization, utilizing participant measurement intervals, age, and gender to standardize normal range variability by age.

The researchers separated the pre-processed dataset into training (80%) and test (20%) datasets, with deep learning performed using simple artificial neural networks (ANN). They examined vital indicators by combining pseudonymized identifying codes and hospitalization dates to generate unique hospitalization identification codes (IHIDs).

They sorted vital sign measurement durations in ascending order and estimated the interval between the vital sign measurements within the IHID.

The researchers identified critical records as data collected six hours before an event, such as a transfer to the ICU or death, and six hours before a CPR occurrence, such as mortality after cardiopulmonary resuscitation.

They divided the records into non-critical and critical categories, eliminating non-critical ones containing crucial events and performing deep learning on the last documented records for IHID in non-critical groups.

The researchers measured the model’s prediction ability using two essential metrics: the area under the receiver-operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC).

Secondary endpoints included CPR, an unexpected transfer to the ICU, and death. The study additionally examined the relevance of the estimation process for all characteristics used and the link between them.

Results

The study included 13,787 individuals with 22,184 hospitalizations and 1,039,070 data points for vital signs. The average participant age at hospitalization was 69 months, with 43% of the patients being female.

The hospitalization lasted 3.0 days. Following data filtering, 14,227 relevant records remained, with 74 months and 43% female.

The critical category accounted for 4.40% of patients, with 261 cases necessitating cardiopulmonary resuscitation, 238 cases involving unscheduled transfer to the ICU, and 141 deaths. The mean imputation value for missing data for the initial measurement interval was 276.

The predictive performance of the generated model was outstanding, with an AUROC of 0.99 and an AUPRC of 0.90.

The team created a deep learning model with exceptional prediction ability that uses simple factors to accurately forecast crucial occurrences while lowering the workload of medical personnel. However, the study was a single-center experiment, warranting further research for external validation of the model.

The most significant predictors of outcomes were the measurement interval, SpO2, and the RR z-score. The model output increases as the interval decreases, whereas the effect decreases as the interval increases. SpO2 showed a similar trend.

Higher respiratory rate and heart rate z-scores had a higher influence on outcomes, whereas lower values of z-scores had less effect.

Examining the association between characteristics for model characterization revealed that narrower measurement intervals resulted in higher SHAP values, but the HR z-score was non-significant.

The association between oxygen saturation (SpO2) values and SHAP readings was continuously inverse, with the tendency becoming more prominent with a decrease in measurement intervals.

Conclusion

Overall, the study findings highlight a model based on deep learning that uses simple data such as vital signs, gender, measurement intervals, and age to predict intervention in failing pediatric patients.

This method decreases medical staff burden by relying on a small number of variables rather than accumulating measurements. The model had AUROC and AUPRC values of 0.99 and 0.90, respectively, much better than earlier research.

The model continuously topped 0.96 for all crucial events, but its AUPRC decreased due to a lack of specialized training. The model produced better findings across all periods, probably due to an imbalance between non-critical and crucial subgroups and data uniformity.