Evaluation of machine learning algorithms for improved risk assessment for Down's syndrome

Area of Science:

• Medical Informatics
• Genetics
• Machine Learning in Healthcare

Background:

• Prenatal screening generates extensive data for disorder risk prediction.
• Current prenatal risk assessment relies on clinical variables and population-specific algorithm optimization.
• First-trimester screening for Down syndrome can be improved with advanced analytical techniques.

Purpose of the Study:

• To evaluate machine learning algorithms for enhancing first-trimester Down syndrome screening performance.
• To develop improved risk assessment models using existing clinical variables.
• To compare machine learning model performance against a commercial risk assessment software.

Main Methods:

• Utilized two real-world datasets for experimenting with multiple classification algorithms.
• Tested implemented models on a third, independent real-world dataset.
• Compared performance metrics (e.g., detection rate, false positive rate, AUC) against a predicate method.

Main Results:

• The best deep neural network model achieved an Area Under the Curve (AUC) of 0.96, with a 78% detection rate and 1% false positive rate.
• A Support Vector Machine (SVM) model yielded an AUC of 0.95, a 61% detection rate, and a 1% false positive rate.
• Optimized deep neural networks outperformed the predicate method, offering higher detection rates at the same false positive rate or similar detection rates with significantly lower false positive rates.

Conclusions:

• Machine learning algorithms offer an adaptive approach to improve prenatal risk assessment models.
• Optimized deep neural networks demonstrate superior performance for first-trimester Down syndrome screening compared to existing methods.
• These findings suggest a potential improvement in screening accuracy by leveraging existing clinical variables and population-specific training data.