Michael Shmoish1, Alina German2,3, Nurit Devir4, Anna Hecht4, Gary Butler5, Aimon Niklasson6, Kerstin Albertsson-Wikland7, Ze'ev Hochberg3. 1. Bioinformatics Knowledge Unit, The Lokey Center, Technion-Israel Institute of Technology, Haifa, Israel. 2. Pediatric Endocrinology, Clalit Health Service, Haifa, Israel. 3. The Ruth and Bruce Rappaport Faculty of Medicine, Technion-Israel Institute of Technology, Haifa, Israel. 4. Computer Science Department, Technion-Israel Institute of Technology, Haifa, Israel. 5. University College London Great Ormond Street Institute of Child Health, London, UK. 6. Göteborg Pediatric Growth Research Center, Department of Pediatrics, Institute of Clinical Sciences, Sahlgrenska Academy, University of Gothenburg, Gothenburg, Sweden. 7. Physiology/Endocrinology, Institute of Neuroscience & Physiology, Sahlgrenska Academy, University of Gothenburg, Gothenburg, Sweden.
Abstract
CONTEXT: Prediction of AH is frequently undertaken in the clinical setting. The commonly used methods are based on the assessment of skeletal maturation. Predictive algorithms generated by machine learning, which can already automatically drive cars and recognize spoken language, are the keys to unlocking data that can precisely inform the pediatrician for real-time decision making. OBJECTIVE: To use machine learning (ML) to predict adult height (AH) based on growth measurements until age 6 years. METHODS: Growth data from 1596 subjects (798 boys) aged 0-20 years from the longitudinal GrowUp 1974 Gothenburg cohort were utilized to train multiple ML regressors. Of these, 100 were used for model comparison, the rest was used for 5-fold cross-validation. The winning model, random forest (RF), was first validated on 684 additional subjects from the 1974 cohort. It was additionally validated using 1890 subjects from the GrowUp 1990 Gothenburg cohort and 145 subjects from the Edinburgh Longitudinal Growth Study cohort. RESULTS: RF with 51 regression trees produced the most accurate predictions. The best predicting features were sex and height at age 3.4-6.0 years. Observed and predicted AHs were 173.9 ± 8.9 cm and 173.9 ± 7.7 cm, respectively, with prediction average error of -0.4 ± 4.0 cm. Validation of prediction for 684 GrowUp 1974 children showed prediction accuracy r = 0.87 between predicted and observed AH (R2 = 0.75). When validated on the 1990 Gothenburg and Edinburgh cohorts (completely unseen by the learned RF model), the prediction accuracy was r = 0.88 in both cases (R2 = 0.77). AH in short children was overpredicted and AH in tall children was underpredicted. Prediction absolute error correlated negatively with AH (P < .0001). CONCLUSION: We show successful, validated ML of AH using growth measurements before age 6 years. The most important features for prediction were sex, and height at age 3.4-6.0. Prediction errors result in over- or underestimates of AH for short and tall subjects, respectively. Prediction by ML can be generalized to other cohorts.
CONTEXT: Prediction of AH is frequently undertaken in the clinical setting. The commonly used methods are based on the assessment of skeletal maturation. Predictive algorithms generated by machine learning, which can already automatically drive cars and recognize spoken language, are the keys to unlocking data that can precisely inform the pediatrician for real-time decision making. OBJECTIVE: To use machine learning (ML) to predict adult height (AH) based on growth measurements until age 6 years. METHODS: Growth data from 1596 subjects (798 boys) aged 0-20 years from the longitudinal GrowUp 1974 Gothenburg cohort were utilized to train multiple ML regressors. Of these, 100 were used for model comparison, the rest was used for 5-fold cross-validation. The winning model, random forest (RF), was first validated on 684 additional subjects from the 1974 cohort. It was additionally validated using 1890 subjects from the GrowUp 1990 Gothenburg cohort and 145 subjects from the Edinburgh Longitudinal Growth Study cohort. RESULTS: RF with 51 regression trees produced the most accurate predictions. The best predicting features were sex and height at age 3.4-6.0 years. Observed and predicted AHs were 173.9 ± 8.9 cm and 173.9 ± 7.7 cm, respectively, with prediction average error of -0.4 ± 4.0 cm. Validation of prediction for 684 GrowUp 1974 children showed prediction accuracy r = 0.87 between predicted and observed AH (R2 = 0.75). When validated on the 1990 Gothenburg and Edinburgh cohorts (completely unseen by the learned RF model), the prediction accuracy was r = 0.88 in both cases (R2 = 0.77). AH in short children was overpredicted and AH in tall children was underpredicted. Prediction absolute error correlated negatively with AH (P < .0001). CONCLUSION: We show successful, validated ML of AH using growth measurements before age 6 years. The most important features for prediction were sex, and height at age 3.4-6.0. Prediction errors result in over- or underestimates of AH for short and tall subjects, respectively. Prediction by ML can be generalized to other cohorts.
Authors: Werner F Blum; Michael B Ranke; Eberhard Keller; Alexandra Keller; Sandra Barth; Christiaan de Bruin; Stefan A Wudy; Jan M Wit Journal: J Endocr Soc Date: 2022-05-06