23/25 ๐๐ฒ๐ป๐๐๐ฒ๐ฃ๐ผ๐๐ฒ: ๐ฃ๐ฎ๐๐ถ๐ฒ๐ป๐-๐๐ฟ๐ฒ๐ฒ, ๐๐ป๐ผ๐๐น๐ฒ๐ฑ๐ด๐ฒ-๐๐ฎ๐๐ฒ๐ฑ ๐ฆ๐ฎ๐ฐ๐ฐ๐ฎ๐ฑ๐ถ๐ฐ ๐๐๐ฒ ๐ ๐ผ๐๐ฒ๐บ๐ฒ๐ป๐ ๐ ๐ผ๐ฑ๐ฒ๐น๐ถ๐ป๐ด ๐ณ๐ผ๐ฟ ๐๐ถ๐ด๐ถ๐๐ฎ๐น ๐ก๐ฒ๐๐ฟ๐ผ๐ฝ๐ต๐๐๐ถ๐ผ๐น๐ผ๐ด๐ถ๐ฐ ๐๐ถ๐ผ๐บ๐ฎ๐ฟ๐ธ๐ฒ๐ฟ ๐๐ฒ๐๐ฒ๐น๐ผ๐ฝ๐บ๐ฒ๐ป๐
To address the lack of robust AI-enabled solutions for detecting saccadic signatures in neurologic diseases due to privacy and scarce datasets, this paper proposes the first fully synthetic, patient-free, multimodal eye movement generation pipeline. A deep learning classifier trained on this synthetic data achieved an AUROC of 0.76 and a sensitivity of 0.71 on real clinical data, demonstrating strong potential for generalizable clinical applications like screening and precise neuroanatomic localization.
#SyntheticData#EyeMovements#SaccadicSignatures#NeurologyAI#DigitalBiomarkers#DeepLearning#MedicalAI#DataGeneration
Paper Link: arxiv.org/abs/2606.09681
H100s were used for most
for datageneration for each iteration, you need around 20 h100 hours
finetuning is possible on colab, and many have been done there itself
usually 3 languages, 3 iterations - full pipeline took ~ 7 days
codebase- github.com/deeps73/CycleDistโฆ
How does @FractionAI_xyz ensure quality data? By having AI agents compete to generate it! This dynamic process, judged by AI trained on human preferences, ensures only the best data defines the next generation of AI. #FractionAI#DataGeneration#QualityData#AIEvolution