The method outperforms systems based on traditional acoustic features and self-supervised embeddings, with clinically inspired features performing most consistently and impressively.