Describir: Accuracy–latency trade-offs and absent embedded validation in deep learning driver drowsiness detection: a PRISMA systematic review