Feature replacement methods enable reliable home video analysis for machine learning detection of autism

Area of Science:

• Neuroscience
• Computer Science
• Developmental Psychology

Background:

• Autism Spectrum Disorder (ASD) affects millions globally, necessitating early diagnosis for effective behavioral therapies.
• Machine learning on video data offers scalable autism detection, but video quality variations introduce missing features, impacting performance.
• Managing missing data is critical for reliable video-based ASD diagnostics.

Purpose of the Study:

• To evaluate the impact of missing values and imputation methods on ASD detection classifiers using video data.
• To compare traditional imputation techniques with novel feature replacement strategies for handling missing video features.

Main Methods:

• Two existing ASD detection classifiers were tested on 140 children's videos.
• Listwise deletion, univariate/multivariate imputation, and novel general/dynamic feature replacement methods were compared.
• Performance was assessed based on ratings of video features and handling of missing data.

Main Results:

• Feature replacement methods (general and dynamic) outperformed classic imputation techniques.
• Algorithmic management of missing values maintained diagnostic fidelity despite variable video quality.
• The study demonstrates the robustness of video-based diagnostics with advanced data imputation.

Conclusions:

• Feature replacement strategies effectively manage missing data in video-based ASD detection.
• Algorithmic approaches can preserve the accuracy of machine learning diagnostics in real-world, variable conditions.
• This research supports the use of video analysis for scalable and reliable early autism screening.