Dataset Information

A practical model for the identification of congenital cataracts using machine learning.

ABSTRACT:

Background

Approximately 1 in 33 newborns is affected by congenital anomalies worldwide. We aimed to develop a practical model for identifying infants with a high risk of congenital cataracts (CCs), which is the leading cause of avoidable childhood blindness.

Methods

This case-control study was performed in the Zhongshan Ophthalmic Center and involved 2005 subjects, including 1274 children with CCs and 731 healthy controls. The CC identification models were established based on birth conditions, family medical history, and family environmental factors using the random forest (RF) and adaptive boosting methods (trained by 1129 CC cases and 609 healthy controls), which were tested by internal 4-fold cross-validation and external validation (145 CC cases and 122 healthy controls). The models were also tested using 4 datasets with gradually reduced proportions of CC patients (bilateral cases) to validate their performance in an approximate simulation of a clinical environment with a relatively low disease prevalence.

Findings

The CC identification models showed high discrimination in both the 4-fold cross validation (area under the curve (AUC)=0.91 [95% confidence interval: 0.88-0.94] in bilateral cases; 0.82 [0.77-0.89] in unilateral cases) and external validation (AUC=0.93±0.05 in bilateral cases; 0.86±0.01 in unilateral cases), and achieved stable performance in the clinical tests (AUC=0.94-0.96 in the four subgroups by RF). Furthermore, family history of CC, low parental education level, and comorbidity were identified as the top three most relevant factors to both bilateral and unilateral CC diagnosis.

Interpretation

Our CC identification models can accurately discriminate CC patients from healthy children and have the potential to serve as a complementary screening procedure, especially in undeveloped and remote areas.

SUBMITTER: Lin D

PROVIDER: S-EPMC6948173 | biostudies-literature |

REPOSITORIES: biostudies-literature

ACCESS DATA

Similar Datasets

Project description:BACKGROUND:Micropeptides are small proteins with length?<?= 100 amino acids. Short open reading frames that could produces micropeptides were traditionally ignored due to technical difficulties, as few small peptides had been experimentally confirmed. In the past decade, a growing number of micropeptides have been shown to play significant roles in vital biological activities. Despite the increased amount of data, we still lack bioinformatics tools for specifically identifying micropeptides from DNA sequences. Indeed, most existing tools for classifying coding and noncoding ORFs were built on datasets in which "normal-sized" proteins were considered to be positives and short ORFs were generally considered to be noncoding. Since the functional and biophysical constraints on small peptides are likely to be different from those on "normal" proteins, methods for predicting short translated ORFs must be trained independently from those for longer proteins. RESULTS:In this study, we have developed MiPepid, a machine-learning tool specifically for the identification of micropeptides. We trained MiPepid using carefully cleaned data from existing databases and used logistic regression with 4-mer features. With only the sequence information of an ORF, MiPepid is able to predict whether it encodes a micropeptide with 96% accuracy on a blind dataset of high-confidence micropeptides, and to correctly classify newly discovered micropeptides not included in either the training or the blind test data. Compared with state-of-the-art coding potential prediction methods, MiPepid performs exceptionally well, as other methods incorrectly classify most bona fide micropeptides as noncoding. MiPepid is alignment-free and runs sufficiently fast for genome-scale analyses. It is easy to use and is available at https://github.com/MindAI/MiPepid. CONCLUSIONS:MiPepid was developed to specifically predict micropeptides, a category of proteins with increasing significance, from DNA sequences. It shows evident advantages over existing coding potential prediction methods on micropeptide identification. It is ready to use and runs fast.

Project description:BackgroundFollowing visible successes on a wide range of predictive tasks, machine learning techniques are attracting substantial interest from medical researchers and clinicians. We address the need for capacity development in this area by providing a conceptual introduction to machine learning alongside a practical guide to developing and evaluating predictive algorithms using freely-available open source software and public domain data.MethodsWe demonstrate the use of machine learning techniques by developing three predictive models for cancer diagnosis using descriptions of nuclei sampled from breast masses. These algorithms include regularized General Linear Model regression (GLMs), Support Vector Machines (SVMs) with a radial basis function kernel, and single-layer Artificial Neural Networks. The publicly-available dataset describing the breast mass samples (N=683) was randomly split into evaluation (n=456) and validation (n=227) samples. We trained algorithms on data from the evaluation sample before they were used to predict the diagnostic outcome in the validation dataset. We compared the predictions made on the validation datasets with the real-world diagnostic decisions to calculate the accuracy, sensitivity, and specificity of the three models. We explored the use of averaging and voting ensembles to improve predictive performance. We provide a step-by-step guide to developing algorithms using the open-source R statistical programming environment.ResultsThe trained algorithms were able to classify cell nuclei with high accuracy (.94 -.96), sensitivity (.97 -.99), and specificity (.85 -.94). Maximum accuracy (.96) and area under the curve (.97) was achieved using the SVM algorithm. Prediction performance increased marginally (accuracy =.97, sensitivity =.99, specificity =.95) when algorithms were arranged into a voting ensemble.ConclusionsWe use a straightforward example to demonstrate the theory and practice of machine learning for clinicians and medical researchers. The principals which we demonstrate here can be readily applied to other complex tasks including natural language processing and image recognition.