Abstract: PO2005
Machine Learning Can Predict the Individual Risk of Acute Pyelonephritis in Children
Session Information
- Pediatric Nephrology: Genetics, Kidney Stones, Quality Improvement, and Case Reports
November 04, 2021 | Location: On-Demand, Virtual Only
Abstract Time: 10:00 AM - 12:00 PM
Category: Pediatric Nephrology
- 1700 Pediatric Nephrology
Authors
- Niel, Olivier, Centre Hospitalier de Luxembourg, Luxembourg, Luxembourg
- François, Pierre, Centre Hospitalier de Luxembourg, Luxembourg, Luxembourg
- Sommelette, Claire, Centre Hospitalier de Luxembourg, Luxembourg, Luxembourg
- Kohnen, Michel, Centre Hospitalier de Luxembourg, Luxembourg, Luxembourg
- Tsobo, Chantal, Centre Hospitalier de Luxembourg, Luxembourg, Luxembourg
- De La Fuente, Isabel, Centre Hospitalier de Luxembourg, Luxembourg, Luxembourg
- Biver, Armand, Centre Hospitalier de Luxembourg, Luxembourg, Luxembourg
Background
Acute pyelonephritis (AP) is a common infection in children. Timely diagnosis of pediatric AP is necessary, since under-diagnosed AP increase infectious morbidity, whereas over-treatment of AP is responsible for an increase in antibiotic resistance and health costs.
However, confirmed AP diagnosis requires validated urine cultures, which can take up to 3 days. Here we propose to use machine learning algorithms to predict the risk of AP in febrile children, using simple parameters available within the first hours of medical care.
Methods
We performed a retrospective study of medical and laboratory files of 102 pediatric patients with a suspected diagnosis of AP, treated between 2014 and 2020 at the pediatric National Reference Hospital of Luxembourg. Based on the results of urine cultures, patients were allocated to the AP or non-AP group. All patients were then randomly split into training and testing batches, used by a Random Forest machine learning algorithm to predict the individual risk of AP, using clinical (age, sex), blood (CRP, white blood cell and neutrophil counts) and urine (red and white blood cell counts) parameters.
Results
Patients’ demographic and clinical characteristics were comparable between groups. In particular, sex ratios were not significantly different between AP and non-AP patients (0.86 versus 0.74).
Random Forest algorithm mean performance metrics were: accuracy 90.48% [85-99%], sensitivity 91.67% [90-95%], specificity 88.89% [80-90%]. Given a prevalence of AP of 60%, positive predictive value was 92.52% [88-95%], negative predictive value 87.67% [82-89%]; mean AUC-ROC was 0.92. Predictions performed with a neural network or a support vector machine algorithm on the same population obtained comparable performance metrics.
Conclusion
Timely diagnosis of pediatric AP is necessary to minimize infectious morbidity, antibiotic resistance and health costs; however, it requires validated urine cultures, which can take several days. Here we showed that machine learning algorithms can accurately predict the individual risk of AP in pediatric patients within the first hours of medical care, helping pediatricians in daily clinical decision making.