A Patch-Based Deep Learning Approach for Detecting Rib Fractures on Frontal Radiographs in Young Children.

Chest radiography is the modality of choice for the identification of rib fractures in young children and there is value for the development of computer-aided rib fracture detection in this age group. However, the automated identification of rib fractures on chest radiographs can be challenging due to the need for high spatial resolution in deep learning frameworks. A patch-based deep learning algorithm was developed to automatically detect rib fractures on frontal chest radiographs in children under 2 years old. A total of 845 chest radiographs of children 0-2 years old (median: 4 months old) were manually segmented for rib fractures by radiologists and served as the ground-truth labels. Image analysis utilized a patch-based sliding-window technique, to meet the high-resolution requirements for fracture detection. Standard transfer learning techniques used ResNet-50 and ResNet-18 architectures. Area-under-curve for precision-recall (AUC-PR) and receiver-operating-characteristic (AUC-ROC), along with patch and whole-image classification metrics, were reported. On the test patches, the ResNet-50 model showed AUC-PR and AUC-ROC of 0.25 and 0.77, respectively, and the ResNet-18 showed an AUC-PR of 0.32 and AUC-ROC of 0.76. On the whole-radiograph level, the ResNet-50 had an AUC-ROC of 0.74 with 88% sensitivity and 43% specificity in identifying rib fractures, and the ResNet-18 had an AUC-ROC of 0.75 with 75% sensitivity and 60% specificity in identifying rib fractures. This work demonstrates the utility of patch-based analysis for detection of rib fractures in children under 2 years old. Future work with large cohorts of multi-institutional data will improve the generalizability of these findings to patients with suspicion of child abuse.

Automating Angle Measurements on Foot Radiographs in Young Children: Feasibility and Performance of a Convolutional Neural Network Model.

Measurement of angles on foot radiographs is an important step in the evaluation of malalignment. The objective is to develop a CNN model to measure angles on radiographs, using radiologists' measurements as the reference standard. This IRB-approved retrospective study included 450 radiographs from 216 patients (< 3 years of age). Angles were automatically measured by means of image segmentation followed by angle calculation, according to Simon's approach for measuring pediatric foot angles. A multiclass U-Net model with a ResNet-34 backbone was used for segmentation. Two pediatric radiologists independently measured anteroposterior and lateral talocalcaneal and talo-1st metatarsal angles using the test dataset and recorded the time used for each study. Intraclass correlation coefficients (ICC) were used to compare angle and paired Wilcoxon signed-rank test to compare time between radiologists and the CNN model. There was high spatial overlap between manual and CNN-based automatic segmentations with dice coefficients ranging between 0.81 (lateral 1st metatarsal) and 0.94 (lateral calcaneus). Agreement was higher for angles on the lateral view when compared to the AP view, between radiologists (ICC: 0.93-0.95, 0.85-0.92, respectively) and between radiologists' mean and CNN calculated (ICC: 0.71-0.73, 0.41-0.52, respectively). Automated angle calculation was significantly faster when compared to radiologists' manual measurements (3 ± 2 vs 114 ± 24 s, respectively; P < 0.001). A CNN model can selectively segment immature ossification centers and automatically calculate angles with a high spatial overlap and moderate to substantial agreement when compared to manual methods, and 39 times faster.

Conceptualizing bias in EHR data: A case study in performance disparities by demographic subgroups for a pediatric obesity incidence classifier.

Electronic Health Records (EHRs) are increasingly used to develop machine learning models in predictive medicine. There has been limited research on utilizing machine learning methods to predict childhood obesity and related disparities in classifier performance among vulnerable patient subpopulations. In this work, classification models are developed to recognize pediatric obesity using temporal condition patterns obtained from patient EHR data in a U.S. study population. We trained four machine learning algorithms (Logistic Regression, Random Forest, Gradient Boosted Trees, and Neural Networks) to classify cases and controls as obesity positive or negative, and optimized hyperparameter settings through a bootstrapping methodology. To assess the classifiers for bias, we studied model performance by population subgroups then used permutation analysis to identify the most predictive features for each model and the demographic characteristics of patients with these features. Mean AUC-ROC values were consistent across classifiers, ranging from 0.72-0.80. Some evidence of bias was identified, although this was through the models performing better for minority subgroups (African Americans and patients enrolled in Medicaid). Permutation analysis revealed that patients from vulnerable population subgroups were over-represented among patients with the most predictive diagnostic patterns. We hypothesize that our models performed better on under-represented groups because the features more strongly associated with obesity were more commonly observed among minority patients. These findings highlight the complex ways that bias may arise in machine learning models and can be incorporated into future research to develop a thorough analytical approach to identify and mitigate bias that may arise from features and within EHR datasets when developing more equitable models.

Deep learning-based prediction of rib fracture presence in frontal radiographs of children under two years of age: a proof-of-concept study.

OBJECTIVE: In this proof-of-concept study, we aimed to develop deep-learning-based classifiers to identify rib fractures on frontal chest radiographs in children under 2 years of age. METHODS: This retrospective study included 1311 frontal chest radiographs (radiographs with rib fractures, n = 653) from 1231 unique patients (median age: 4 m). Patients with more than one radiograph were included only in the training set. A binary classification was performed to identify the presence or absence of rib fractures using transfer learning and Resnet-50 and DenseNet-121 architectures. The area under the receiver operating characteristic curve (AUC-ROC) was reported. Gradient-weighted class activation mapping was used to highlight the region most relevant to the deep learning models' predictions. RESULTS: On the validation set, the ResNet-50 and DenseNet-121 models obtained an AUC-ROC of 0.89 and 0.88, respectively. On the test set, the ResNet-50 model demonstrated an AUC-ROC of 0.84 with a sensitivity of 81% and specificity of 70%. The DenseNet-50 model obtained an AUC of 0.82 with 72% sensitivity and 79% specificity. CONCLUSION: In this proof-of-concept study, a deep learning-based approach enabled the automatic detection of rib fractures in chest radiographs of young children with performances comparable to pediatric radiologists. Further evaluation of this approach on large multi-institutional data sets is needed to assess the generalizability of our results. ADVANCES IN KNOWLEDGE: In this proof-of-concept study, a deep learning-based approach performed well in identifying chest radiographs with rib fractures. These findings provide further impetus to develop deep learning algorithms for identifying rib fractures in children, especially those with suspected physical abuse or non-accidental trauma.

Personalized prediction of early childhood asthma persistence: A machine learning approach.

Early childhood asthma diagnosis is common; however, many children diagnosed before age 5 experience symptom resolution and it remains difficult to identify individuals whose symptoms will persist. Our objective was to develop machine learning models to identify which individuals diagnosed with asthma before age 5 continue to experience asthma-related visits. We curated a retrospective dataset for 9,934 children derived from electronic health record (EHR) data. We trained five machine learning models to differentiate individuals without subsequent asthma-related visits (transient diagnosis) from those with asthma-related visits between ages 5 and 10 (persistent diagnosis) given clinical information up to age 5 years. Based on average NPV-Specificity area (ANSA), all models performed significantly better than random chance, with XGBoost obtaining the best performance (0.43 mean ANSA). Feature importance analysis indicated age of last asthma diagnosis under 5 years, total number of asthma related visits, self-identified black race, allergic rhinitis, and eczema as important features. Although our models appear to perform well, a lack of prior models utilizing a large number of features to predict individual persistence makes direct comparison infeasible. However, feature importance analysis indicates our models are consistent with prior research indicating diagnosis age and prior health service utilization as important predictors of persistent asthma. We therefore find that machine learning models can predict which individuals will experience persistent asthma with good performance and may be useful to guide clinician and parental decisions regarding asthma counselling in early childhood.