Using Deep Learning to Automate Eosinophil Counting in Pediatric Ulcerative Colitis Histopathological Images.

Background: Accurate identification of inflammatory cells from mucosal histopathology images is important in diagnosing ulcerative colitis. The identification of eosinophils in the colonic mucosa has been associated with disease course. Cell counting is not only time-consuming but can also be subjective to human biases. In this study we developed an automatic eosinophilic cell counting tool from mucosal histopathology images, using deep learning. Method: Four pediatric IBD pathologists from two North American pediatric hospitals annotated 530 crops from 143 standard-of-care hematoxylin and eosin (H & E) rectal mucosal biopsies. A 305/75 split was used for training/validation to develop and optimize a U-Net based deep learning model, and 150 crops were used as a test set. The U-Net model was then compared to SAU-Net, a state-of-the-art U-Net variant. We undertook post-processing steps, namely, (1) the pixel-level probability threshold, (2) the minimum number of clustered pixels to designate a cell, and (3) the connectivity. Experiments were run to optimize model parameters using AUROC and cross-entropy loss as the performance metrics. Results: The F1-score was 0.86 (95%CI:0.79-0.91) (Precision: 0.77 (95%CI:0.70-0.83), Recall: 0.96 (95%CI:0.93-0.99)) to identify eosinophils as compared to an F1-score of 0.2 (95%CI:0.13-0.26) for SAU-Net (Precision: 0.38 (95%CI:0.31-0.46), Recall: 0.13 (95%CI:0.08-0.19)). The inter-rater reliability was 0.96 (95%CI:0.93-0.97). The correlation between two pathologists and the algorithm was 0.89 (95%CI:0.82-0.94) and 0.88 (95%CI:0.80-0.94) respectively. Conclusion: Our results indicate that deep learning-based automated eosinophilic cell counting can obtain a robust level of accuracy with a high degree of concordance with manual expert annotations.

Deep-learning computer vision can identify increased nuchal translucency in the first trimester of pregnancy.

OBJECTIVE: Many fetal anomalies can already be diagnosed by ultrasound in the first trimester of pregnancy. Unfortunately, in clinical practice, detection rates for anomalies in early pregnancy remain low. Our aim was to use an automated image segmentation algorithm to detect one of the most common fetal anomalies: a thickened nuchal translucency (NT), which is a marker for genetic and structural anomalies. METHODS: Standardized mid-sagittal ultrasound images of the fetal head and chest were collected for 560 fetuses between 11 and 13 weeks and 6 days of gestation, 88 (15.7%) of whom had an NT thicker than 3.5 mm. Image quality was graded as high or low by two fetal medicine experts. Images were divided into a training-set (n = 451, 55 thick NT) and a test-set (n = 109, 33 thick NT). We then trained a U-Net convolutional neural network to segment the fetus and the NT region and computed the NT:fetus ratio of these regions. The ability of this ratio to separate thick (anomalous) NT regions from healthy, typical NT regions was first evaluated in ground-truth segmentation to validate the metric and then with predicted segmentation to validate our algorithm, both using the area under the receiver operator curve (AUROC). RESULTS: The ground-truth NT:fetus ratio detected thick NTs with 0.97 AUROC in both the training and test sets. The fetus and NT regions were detected with a Dice score of 0.94 in the test set. The NT:fetus ratio based on model segmentation detected thick NTs with an AUROC of 0.96 relative to clinician labels. At a 91% specificity, 94% of thick NT cases were detected (sensitivity) in the test set. The detection rate was statistically higher (p = 0.003) in high versus low-quality images (AUROC 0.98 vs. 0.90, respectively). CONCLUSION: Our model provides an explainable deep-learning method for detecting increased NT. This technique can be used to screen for other fetal anomalies in the first trimester of pregnancy.

Application of STREAM-URO and APPRAISE-AI reporting standards for artificial intelligence studies in pediatric urology: A case example with pediatric hydronephrosis.

INTRODUCTION: Artificial intelligence (AI) and machine learning (ML) in pediatric urology is gaining increased popularity and credibility. However, the literature lacks standardization in reporting and there are areas for methodological improvement, which incurs difficulty in comparison between studies and may ultimately hurt clinical implementation of these models. The "STandardized REporting of Applications of Machine learning in UROlogy" (STREAM-URO) framework provides methodological instructions to improve transparent reporting in urology and APPRAISE-AI in a critical appraisal tool which provides quantitative measures for the quality of AI studies. The adoption of these will allow urologists and developers to ensure consistency in reporting, improve comparison, develop better models, and hopefully inspire clinical translation. METHODS: In this article, we have applied STREAM-URO framework and APPRAISE-AI tool to the pediatric hydronephrosis literature. By doing this, we aim to describe best practices on ML reporting in urology with STREAM-URO and provide readers with a critical appraisal tool for ML quality with APPRAISE-AI. By applying these to the pediatric hydronephrosis literature, we provide some tutorial for other readers to employ these in developing and appraising ML models. We also present itemized recommendations for adequate reporting, and critically appraise the quality of ML in pediatric hydronephrosis insofar. We provide examples of strong reporting and highlight areas for improvement. RESULTS: There were 8 ML models applied to pediatric hydronephrosis. The 26-item STREAM-URO framework is provided in Appendix A and 24-item APPRAISE-AI tool is provided in Appendix B. Across the 8 studies, the median compliance with STREAM-URO was 67 % and overall study quality was moderate. The highest scoring APPRAISE-AI domains in pediatric hydronephrosis were clinical relevance and reporting quality, while the worst were methodological conduct, robustness of results, and reproducibility. CONCLUSIONS: If properly conducted and reported, ML has the potential to impact the care we provide to patients in pediatric urology. While AI is exciting, the paucity of strong evidence limits our ability to translate models to practice. The first step toward this goal is adequate reporting and ensuring high quality models, and STREAM-URO and APPRAISE-AI can facilitate better reporting and critical appraisal, respectively.

Early prediction of pediatric asthma in the Canadian Healthy Infant Longitudinal Development (CHILD) birth cohort using machine learning.

BACKGROUND: Early identification of children at risk of asthma can have significant clinical implications for effective intervention and treatment. This study aims to disentangle the relative timing and importance of early markers of asthma. METHODS: Using the CHILD Cohort Study, 132 variables measured in 1754 multi-ethnic children were included in the analysis for asthma prediction. Data up to 4 years of age was used in multiple machine learning models to predict physician-diagnosed asthma at age 5 years. Both predictive performance and variable importance was assessed in these models. RESULTS: Early-life data (≤1 year) has limited predictive ability for physician-diagnosed asthma at age 5 years (area under the precision-recall curve (AUPRC) < 0.35). The earliest reliable prediction of asthma is achieved at age 3 years, (area under the receiver-operator curve (AUROC) > 0.90) and (AUPRC > 0.80). Maternal asthma, antibiotic exposure, and lower respiratory tract infections remained highly predictive throughout childhood. Wheezing status and atopy are the most important predictors of early childhood asthma from among the factors included in this study. CONCLUSIONS: Childhood asthma is predictable from non-biological measurements from the age of 3 years, primarily using parental asthma and patient history of wheezing, atopy, antibiotic exposure, and lower respiratory tract infections. IMPACT: Machine learning models can predict physician-diagnosed asthma in early childhood (AUROC > 0.90 and AUPRC > 0.80) using ≥3 years of non-biological and non-genetic information, whereas prediction with the same patient information available before 1 year of age is challenging. Wheezing, atopy, antibiotic exposure, lower respiratory tract infections, and the child's mother having asthma were the strongest early markers of 5-year asthma diagnosis, suggesting an opportunity for earlier diagnosis and intervention and focused assessment of patients at risk for asthma, with an evolving risk stratification over time.

Predicting obstruction risk using common ultrasonography parameters in paediatric hydronephrosis with machine learning.

OBJECTIVE: To sensitively predict the risk of renal obstruction on diuretic renography using routine reported ultrasonography (US) findings, coupled with machine learning approaches, and determine safe criteria for deferral of diuretic renography. PATIENTS AND METHODS: Patients from two institutions with isolated hydronephrosis who underwent a diuretic renogram within 3 months following renal US were included. Age, sex, and routinely reported US findings (laterality, kidney length, anteroposterior diameter, Society for Fetal Urology [SFU] grade) were abstracted. The drainage half-times were collected from renography and stratified as low risk (<20 min, primary outcome), intermediate risk (20-60 min), and high risk of obstruction (>60 min). A random Forest model was trained to classify obstruction risk, here named the 'Artificial intelligence Evaluation of Renogram Obstruction' (AERO). Model performance was determined by measuring area under the receiver-operating-characteristic curve (AUROC) and decision curve analysis. RESULTS: A total of 304 patients met the inclusion criteria, with a median (interquartile range) age of diuretic renogram at 4 (2-7) months. Of all patients, 48 (16%) were low risk, 102 (33%) were intermediate risk, 156 (51%) were high risk of obstruction based on diuretic renogram. The AERO achieved a binary AUROC of 0.84, multi-class AUROC of 0.74 that was superior to the SFU grade, and external validation (n = 64) binary AUROC of 0.76. The most important features for prediction included age, anteroposterior diameter, and SFU grade. We deployed our application in an easy-to-use application (https://sickkidsurology.shinyapps.io/AERO/). At a threshold probability of 30%, the AERO would allow 66 more patients per 1000 to safely avoid a renogram without missing significant obstruction compared to a strategy in which a renogram is routinely performed for SFU Grade ≥3. CONCLUSIONS: Coupled with machine learning, routine US findings can improve the criteria to determine in which children with isolated hydronephrosis a diuretic renogram can be safely avoided. Further optimisation and validation are required prior to implementation into clinical practice.

Trends in Management of Fetuses with Suspected Lower Urinary Tract Obstruction (LUTO): A High-Risk Fetal and Pediatric Center Experience in a Universal-Access-to-Care System.

INTRODUCTION: Neonates with lower urinary tract obstruction (LUTO) experience high morbidity and mortality associated with the development of chronic kidney disease. The prenatal detection rate for LUTO is less than 50%, with late or missed diagnosis leading to delayed management and long-term sequelae in the remainder. We aimed to explore the trends in prenatal detection and management at a high-risk fetal center and determine if similar trends of postnatal presentations were noted for the same period. METHODS: Prenatal and postnatal LUTO databases from a tertiary fetal center and its associated pediatric center between 2009 and 2021 were reviewed, capturing maternal age, gestational age (GA) at diagnosis, and rates of termination of pregnancy (TOP). Time series analysis using autocorrelation was performed to investigate time trend changes for prenatally suspected and postnatally confirmed LUTO cases. RESULTS: A total of 161 fetuses with prenatally suspected LUTO were identified, including 78 terminations. No significant time trend was found when evaluating the correlation between time periods, prenatal suspicion, and postnatal confirmation of LUTO cases (Durbin-Watson [DW] = 1.99, p = 0.3641 and DW = 2.86, p = 0.9113, respectively). GA at referral was 20.0 weeks (interquartile range [IQR] 12, 35) and 22.0 weeks (IQR 13, 37) for TOP and continued pregnancies (p < 0.0001). GA at initial ultrasound was earlier in terminated fetuses compared to continued (20.0 [IQR 12, 35] weeks vs. 22.5 [IQR 13, 39] weeks, p < 0.0001). While prenatal LUTO suspicion remained consistently higher than postnatal presentations, the rates of postnatal presentations and terminations remained stable during the study years (p = 0.7913 and 0.2338), as were GA at TOP and maternal age at diagnosis (p = 0.1710 and 0.1921). CONCLUSION: This study demonstrated that more severe cases of LUTO are referred earlier and are more likely to undergo TOP. No significant trend was detected between time and prenatally suspected or postnatally confirmed LUTO, highlighting the need for further studies to better delineate factors that can increase prenatal detection.

Automatic Prediction of Paediatric Cardiac Output From Echocardiograms Using Deep Learning Models.

Background: Cardiac output (CO) perturbations are common and cause significant morbidity and mortality. Accurate CO assessment is crucial for guiding treatment in anaesthesia and critical care, but measurement is difficult, even for experts. Artificial intelligence methods show promise as alternatives for accurate, rapid CO assessment. Methods: We reviewed paediatric echocardiograms with normal CO and a dilated cardiomyopathy patient group with reduced CO. Experts measured the left ventricular outflow tract diameter, velocity time integral, CO, and cardiac index (CI). EchoNet-Dynamic is a deep learning model for estimation of ejection fraction in adults. We modified this model to predict the left ventricular outflow tract diameter and retrained it on paediatric data. We developed a novel deep learning approach for velocity time integral estimation. The combined models enable automatic prediction of CO. We evaluated the models against expert measurements. Primary outcomes were root-mean-squared error, mean absolute error, mean average percentage error, and coefficient of determination (R2). Results: In a test set unused during training, CI was estimated with the root-mean-squared error of 0.389 L/min/m2, mean absolute error of 0.321 L/min/m2, mean average percentage error of 10.8%, and R2 of 0.755. The Bland-Altman analysis showed that the models estimated CI with a bias of +0.14 L/min/m2 and 95% limits of agreement -0.58 to 0.86 L/min/m2. Conclusions: Our model estimated CO with strong correlation to ground truth and a bias of 0.17 L/min, better than many CO measurements in paediatrics. Model pretraining enabled accurate estimation despite a small dataset. Potential uses include supporting clinicians in real-time bedside calculation of CO, identification of low-CO states, and treatment responses.

Contexte: Les perturbations du débit cardiaque sont fréquentes et associées à des taux élevés de morbidité et de mortalité. Une évaluation juste du débit cardiaque est essentielle pour orienter le choix du traitement anesthésique et des soins critiques. Or, il est difficile de mesurer le débit cardiaque, même pour les experts. Les méthodes fondées sur l'intelligence artificielle semblent toutefois prometteuses pour évaluer le débit cardiaque avec exactitude et rapidité. Méthodologie: Nous avons analysé des échocardiogrammes pédiatriques chez des personnes dont le débit cardiaque est normal ainsi que chez des patients qui étaient atteints d'une cardiomyopathie dilatée et dont le débit cardiaque était réduit. Des experts ont mesuré le diamètre de la voie d'éjection ventriculaire gauche, l'intégrale de la vitesse par rapport au temps (IVT), le débit cardiaque et l'index cardiaque. L'outil EchoNet-Dynamic est un modèle d'apprentissage profond qui donne une estimation de la fraction d'éjection chez les adultes. Nous avons modifié ce modèle afin qu'il puisse prédire le diamètre de la voie d'éjection ventriculaire gauche et l'avons entraîné à l'aide de données pédiatriques. Nous avons également mis au point une nouvelle approche d'apprentissage profond pour l'estimation des valeurs d'IVT. La combinaison de ces modèles a permis de prédire de façon automatique le débit cardiaque, et nous avons évalué les résultats obtenus par rapport à ceux des experts. Les principaux critères d'évaluation étaient l'erreur moyenne quadratique (EMQ), l'erreur moyenne absolue (EMA), le pourcentage d'erreur moyen (PEM) ainsi que le coefficient de détermination (R2). Résultats: Dans un ensemble d'essais n'ayant pas été utilisé au cours de l'entraînement du modèle, l'index cardiaque a été estimé avec une EMQ de 0,389 L/min/m2, une EMA de 0,321 L/min/m2, un PEM de 10,8 % et un R2 de 0,755. Selon l'analyse de Bland-Altman, le biais pour les estimations de l'index cardiaque était de + 0,14 L/min/m2, et les limites de concordance à 95 % étaient de ­0,58 à 0,86 L/min/m2. Conclusions: Les estimations générées par le modèle pour le débit cardiaque montraient une forte corrélation avec les valeurs de référence et un biais à 0,17 L/min, ce qui est mieux que bien des mesures du débit cardiaque utilisées en pédiatrie. Malgré un petit ensemble de données, le modèle entraîné a permis de produire une estimation juste. Les utilisations potentielles comprennent l'aide aux cliniciens dans le calcul du débit cardiaque en temps réel et au chevet du patient, le dépistage d'un faible débit cardiaque et l'évaluation de la réponse au traitement.

Dissociable default-mode subnetworks subserve childhood attention and cognitive flexibility: Evidence from deep learning and stereotactic electroencephalography.

Cognitive flexibility encompasses the ability to efficiently shift focus and forms a critical component of goal-directed attention. The neural substrates of this process are incompletely understood in part due to difficulties in sampling the involved circuitry. We leverage stereotactic intracranial recordings to directly resolve local-field potentials from otherwise inaccessible structures to study moment-to-moment attentional activity in children with epilepsy performing a flexible attentional task. On an individual subject level, we employed deep learning to decode neural features predictive of task performance indexed by single-trial reaction time. These models were subsequently aggregated across participants to identify predictive brain regions based on AAL atlas and FIND functional network parcellations. Through this approach, we show that fluctuations in beta (12-30 Hz) and gamma (30-80 Hz) power reflective of increased top-down attentional control and local neuronal processing within relevant large-scale networks can accurately predict single-trial task performance. We next performed connectomic profiling of these highly predictive nodes to examine task-related engagement of distributed functional networks, revealing exclusive recruitment of the dorsal default mode network during shifts in attention. The identification of distinct substreams within the default mode system supports a key role for this network in cognitive flexibility and attention in children. Furthermore, convergence of our results onto consistent functional networks despite significant inter-subject variability in electrode implantations supports a broader role for deep learning applied to intracranial electrodes in the study of human attention.

A quantitative analysis of voiding cystourethrogram features confirms the association between high-grade vesicoureteral reflux with male sex, younger age, and hydronephrosis.

INTRODUCTION: Vesicoureteral reflux (VUR) is commonly diagnosed in the workup of urinary tract infections or hydronephrosis in children. Traditionally, VUR severity is graded subjectively based on voiding cystourethrogram (VCUG) imaging. Herein, we characterized the association between age, sex, and indication for VCUG, by employing standardized quantitative features. METHODS: We included renal units with a high certainty in VUR grade (>80% consensus) from the qVUR model validation study at our institution between 2013 and 2019. We abstracted the following variables: age, sex, laterality, indication for VCUG, and qVUR parameters (tortuosity, ureter widths on VCUG). High-grade VUR was defined as grade 4 or 5 The association between each variable and VUR grade was assessed. RESULTS: A total of 443 patients (523 renal units) were included, consisting of a 48:52 male/female ratio. The median age at VCUG was 13 months. Younger age at VCUG (<6 months) was associated with greater odds of severe VUR (odds ratio [OR] 2.0), and there was a weak correlation between age and VUR grade (ρ=-0.17). Male sex was associated with increased odds of high-grade VUR (OR 2.7). VCUGs indicated for hydronephrosis were associated with high-grade VUR (OR 4.1) compared to those indicated for UTI only. Ureter tortuosity and width were significantly associated with each clinical variable and VUR severity. CONCLUSIONS: Male sex, younger age (<6 months), and history of hydronephrosis are associated with both high-grade VUR and standardized quantitative measures, including greater ureter tortuosity and increased ureteral width. This lends support to quantitative assessment to improve reliability in VUR grading.

Machine learning-based mortality prediction models using national liver transplantation registries are feasible but have limited utility across countries.

Many countries curate national registries of liver transplant (LT) data. These registries are often used to generate predictive models; however, potential performance and transferability of these models remain unclear. We used data from 3 national registries and developed machine learning algorithm (MLA)-based models to predict 90-day post-LT mortality within and across countries. Predictive performance and external validity of each model were assessed. Prospectively collected data of adult patients (aged ≥18 years) who underwent primary LTs between January 2008 and December 2018 from the Canadian Organ Replacement Registry (Canada), National Health Service Blood and Transplantation (United Kingdom), and United Network for Organ Sharing (United States) were used to develop MLA models to predict 90-day post-LT mortality. Models were developed using each registry individually (based on variables inherent to the individual databases) and using all 3 registries combined (variables in common between the registries [harmonized]). The model performance was evaluated using area under the receiver operating characteristic (AUROC) curve. The number of patients included was as follows: Canada, n = 1214; the United Kingdom, n = 5287; and the United States, n = 59,558. The best performing MLA-based model was ridge regression across both individual registries and harmonized data sets. Model performance diminished from individualized to the harmonized registries, especially in Canada (individualized ridge: AUROC, 0.74; range, 0.73-0.74; harmonized: AUROC, 0.68; range, 0.50-0.73) and US (individualized ridge: AUROC, 0.71; range, 0.70-0.71; harmonized: AUROC, 0.66; range, 0.66-0.66) data sets. External model performance across countries was poor overall. MLA-based models yield a fair discriminatory potential when used within individual databases. However, the external validity of these models is poor when applied across countries. Standardization of registry-based variables could facilitate the added value of MLA-based models in informing decision making in future LTs.

Differential Expression Enrichment Tool (DEET): an interactive atlas of human differential gene expression.

Differential gene expression analysis using RNA sequencing (RNA-seq) data is a standard approach for making biological discoveries. Ongoing large-scale efforts to process and normalize publicly available gene expression data enable rapid and systematic reanalysis. While several powerful tools systematically process RNA-seq data, enabling their reanalysis, few resources systematically recompute differentially expressed genes (DEGs) generated from individual studies. We developed a robust differential expression analysis pipeline to recompute 3162 human DEG lists from The Cancer Genome Atlas, Genotype-Tissue Expression Consortium, and 142 studies within the Sequence Read Archive. After measuring the accuracy of the recomputed DEG lists, we built the Differential Expression Enrichment Tool (DEET), which enables users to interact with the recomputed DEG lists. DEET, available through CRAN and RShiny, systematically queries which of the recomputed DEG lists share similar genes, pathways, and TF targets to their own gene lists. DEET identifies relevant studies based on shared results with the user's gene lists, aiding in hypothesis generation and data-driven literature review.

Using a machine learning algorithm to predict outcome of primary cytoreductive surgery in advanced ovarian cancer.

OBJECTIVE: To develop a machine learning (ML) algorithm to predict outcome of primary cytoreductive surgery (PCS) in patients with advanced ovarian cancer (AOC) METHODS: This retrospective cohort study included patients with AOC undergoing PCS between January 2017 and February 2021. Using radiologic criteria, patient factors (age, CA-125, performance status, BRCA) and surgical complexity scores, we trained a random forest model to predict the dichotomous outcome of optimal cytoreduction (<1 cm) and no gross residual (RD = 0 mm) using JMP-Pro 15 (SAS). This model is available at https://ipm-ml.ccm.sickkids.ca. RESULTS: One hundred and fifty-one patients underwent PCS and randomly assigned to train (n = 92), validate (n = 30), or test (n = 29) the model. The median age was 58 (27-83). Patients with suboptimal cytoreduction were more likely to have an Eastern Cooperative Oncology Group 3-4 (11% vs. 0.75%, p = 0.004), lower albumin (38 vs. 41, p = 0.02), and higher CA125 (1126 vs. 388, p = 0.012) than patients with optimal cytoreduction (n = 133). There were no significant differences in age, histology, stage, or BRCA status between groups. The bootstrap random forest model had AUCs of 99.8% (training), 89.6%(validation), and 89.0% (test). The top five contributors were CA125, albumin, diaphragmatic disease, age, and ascites. For RD = 0 mm, the AUCs were 94.4%, 52%, and 84%, respectively. CONCLUSION: Our ML algorithm demonstrated high accuracy in predicting optimal cytoreduction in patients with AOC selected for PCS and may assist decision-making.

Systemic lupus erythematosus phenotypes formed from machine learning with a specific focus on cognitive impairment.

OBJECTIVE: To phenotype SLE based on symptom burden (disease damage, system involvement and patient reported outcomes), with a specific focus on objective and subjective cognitive function. METHODS: SLE patients ages 18-65 years underwent objective cognitive assessment using the ACR Neuropsychological Battery (ACR-NB) and data were collected on demographic and clinical variables, disease burden/activity, health-related quality of life (HRQoL), depression, anxiety, fatigue and perceived cognitive deficits. Similarity network fusion (SNF) was used to identify patient subtypes. Differences between the subtypes were evaluated using Kruskal-Wallis and χ2 tests. RESULTS: Of the 238 patients, 90% were female, with a mean age of 41 years (s.d. 12) and a disease duration of 14 years (s.d. 10) at the study visit. The SNF analysis defined two subtypes (A and B) with distinct patterns in objective and subjective cognitive function, disease burden/damage, HRQoL, anxiety and depression. Subtype A performed worst on all significantly different tests of objective cognitive function (P < 0.03) compared with subtype B. Subtype A also had greater levels of subjective cognitive function (P < 0.001), disease burden/damage (P < 0.04), HRQoL (P < 0.001) and psychiatric measures (P < 0.001) compared with subtype B. CONCLUSION: This study demonstrates the complexity of cognitive impairment (CI) in SLE and that individual, multifactorial phenotypes exist. Those with greater disease burden, from SLE-specific factors or other factors associated with chronic conditions, report poorer cognitive functioning and perform worse on objective cognitive measures. By exploring different ways of phenotyping SLE we may better define CI in SLE. Ultimately this will aid our understanding of personalized CI trajectories and identification of appropriate treatments.

Deep learning imaging features derived from kidney ultrasounds predict chronic kidney disease progression in children with posterior urethral valves.

BACKGROUND: We sought to use deep learning to extract anatomic features from postnatal kidney ultrasounds and evaluate their performance in predicting the risk and timing of chronic kidney disease (CKD) progression for boys with posterior urethral valves (PUV). We hypothesized that these features would predict CKD progression better than clinical characteristics such as nadir creatinine alone. METHODS: We performed a retrospective cohort study of boys with PUV treated at two pediatric health systems from 1990 to 2021. Features of kidneys were extracted from initial postnatal kidney ultrasound images using a deep learning model. Three time-to-event prediction models were built using random survival forests. The Imaging Model included deep learning imaging features, the Clinical Model included clinical data, and the Ensemble Model combined imaging features and clinical data. Separate models were built to include time-dependent clinical data that were available at 6 months, 1 year, 3 years, and 5 years. RESULTS: Two-hundred and twenty-five patients were included in the analysis. All models performed well with C-indices of 0.7 or greater. The Clinical Model outperformed the Imaging Model at all time points with nadir creatinine driving the performance of the Clinical Model. Combining the 6-month Imaging Model (C-index 0.7; 95% confidence interval [CI] 0.6, 0.79) with the 6-month Clinical Model (C-index 0.79; 95% CI 0.71, 0.86) resulted in a 6-month Ensemble Model that performed better (C-index 0.82; 95% CI 0.77, 0.88) than either model alone. CONCLUSIONS: Deep learning imaging features extracted from initial postnatal kidney ultrasounds may improve early prediction of CKD progression among children with PUV. A higher resolution version of the Graphical abstract is available as Supplementary information.

Multi-institutional Validation of Improved Vesicoureteral Reflux Assessment With Simple and Machine Learning Approaches.

PURPOSE: Vesicoureteral reflux grading from voiding cystourethrograms is highly subjective with low reliability. We aimed to demonstrate improved reliability for vesicoureteral reflux grading with simple and machine learning approaches using ureteral tortuosity and dilatation on voiding cystourethrograms. MATERIALS AND METHODS: Voiding cystourethrograms were collected from our institution for training and 5 external data sets for validation. Each voiding cystourethrogram was graded by 5-7 raters to determine a consensus vesicoureteral reflux grade label and inter- and intra-rater reliability was assessed. Each voiding cystourethrogram was assessed for 4 features: ureteral tortuosity, proximal, distal, and maximum ureteral dilatation. The labels were then assigned to the combination of the 4 features. A machine learning-based model, qVUR, was trained to predict vesicoureteral reflux grade from these features and model performance was assessed by AUROC (area under the receiver-operator-characteristic). RESULTS: A total of 1,492 kidneys and ureters were collected from voiding cystourethrograms resulting in a total of 8,230 independent gradings. The internal inter-rater reliability for vesicoureteral reflux grading was 0.44 with a median percent agreement of 0.71 and low intra-rater reliability. Higher values for each feature were associated with higher vesicoureteral reflux grade. qVUR performed with an accuracy of 0.62 (AUROC=0.84) with stable performance across all external data sets. The model improved vesicoureteral reflux grade reliability by 3.6-fold compared to traditional grading (P < .001). CONCLUSIONS: In a large pediatric population from multiple institutions, we show that machine learning-based assessment for vesicoureteral reflux improves reliability compared to current grading methods. qVUR is generalizable and robust with similar accuracy to clinicians but the added prognostic value of quantitative measures warrants further study.

The silent trial - the bridge between bench-to-bedside clinical AI applications.

As more artificial intelligence (AI) applications are integrated into healthcare, there is an urgent need for standardization and quality-control measures to ensure a safe and successful transition of these novel tools into clinical practice. We describe the role of the silent trial, which evaluates an AI model on prospective patients in real-time, while the end-users (i.e., clinicians) are blinded to predictions such that they do not influence clinical decision-making. We present our experience in evaluating a previously developed AI model to predict obstructive hydronephrosis in infants using the silent trial. Although the initial model performed poorly on the silent trial dataset (AUC 0.90 to 0.50), the model was refined by exploring issues related to dataset drift, bias, feasibility, and stakeholder attitudes. Specifically, we found a shift in distribution of age, laterality of obstructed kidneys, and change in imaging format. After correction of these issues, model performance improved and remained robust across two independent silent trial datasets (AUC 0.85-0.91). Furthermore, a gap in patient knowledge on how the AI model would be used to augment their care was identified. These concerns helped inform the patient-centered design for the user-interface of the final AI model. Overall, the silent trial serves as an essential bridge between initial model development and clinical trials assessment to evaluate the safety, reliability, and feasibility of the AI model in a minimal risk environment. Future clinical AI applications should make efforts to incorporate this important step prior to embarking on a full-scale clinical trial.

Machine learning classification of multiple sclerosis in children using optical coherence tomography.

BACKGROUND: In children, multiple sclerosis (MS) is the ultimate diagnosis in only 1/5 to 1/3 of cases after a first episode of central nervous system (CNS) demyelination. As the visual pathway is frequently affected in MS and other CNS demyelinating disorders (DDs), structural retinal imaging such as optical coherence tomography (OCT) can be used to differentiate MS. OBJECTIVE: This study aimed to investigate the utility of machine learning (ML) based on OCT features to identify distinct structural retinal features in children with DDs. METHODS: This study included 512 eyes from 187 (neyes = 374) children with demyelinating diseases and 69 (neyes = 138) controls. Input features of the analysis comprised of 24 auto-segmented OCT features. RESULTS: Random Forest classifier with recursive feature elimination yielded the highest predictive values and identified DDs with 75% and MS with 80% accuracy, while multiclass distinction between MS and monophasic DD was performed with 64% accuracy. A set of eight retinal features were identified as the most important features in this classification. CONCLUSION: This study demonstrates that ML based on OCT features can be used to support a diagnosis of MS in children.

Data governance functions to support responsible data stewardship in pediatric radiology research studies using artificial intelligence.

The integration of human and machine intelligence promises to profoundly change the practice of medicine. The rapidly increasing adoption of artificial intelligence (AI) solutions highlights its potential to streamline physician work and optimize clinical decision-making, also in the field of pediatric radiology. Large imaging databases are necessary for training, validating and testing these algorithms. To better promote data accessibility in multi-institutional AI-enabled radiologic research, these databases centralize the large volumes of data required to effect accurate models and outcome predictions. However, such undertakings must consider the sensitivity of patient information and therefore utilize requisite data governance measures to safeguard data privacy and security, to recognize and mitigate the effects of bias and to promote ethical use. In this article we define data stewardship and data governance, review their key considerations and applicability to radiologic research in the pediatric context, and consider the associated best practices along with the ramifications of poorly executed data governance. We summarize several adaptable data governance frameworks and describe strategies for their implementation in the form of distributed and centralized approaches to data management.