NarraSpace
A multimodal interpretability framework for early screening of Alzheimer's disease. By linking transcripts of a picture-description task with visual scene elements and predefined scene events, it constructs traceable classification evidence, improving model credibility in clinical review.
Timeline (last 90 days)
The NarraSpace framework links words in the transcript to named elements in the image and retrieves utterances related to predefined scene events, in order to construct traceable classification evidence.
NarraSpace achieves an average accuracy of 80.88%±2.56% across 50 random seeds on the official 48-person test set of the ADReSS benchmark.
After averaging probabilities, NarraSpace achieves 81.25% accuracy and an AUC of 91.32% on the ADReSS test set.
NarraSpace's classification stage uses a frozen language model to compare text for each scene event, combined with transcript surprisal and coverage measurements, and finally predicts with a fixed SVM classifier.
NarraSpace retains source records, and every event score can be linked back to its source utterance, speech interval, and lexical match record.
NarraSpace's evaluation is based on the 48-person test set of the ADReSS benchmark, so the sample size is relatively limited, and the coverage of predefined scene events may also be constrained by the design of the image itself.
All Facts (6)
The NarraSpace framework links words in the transcript to named elements in the image and retrieves utterances related to predefined scene events, in order to construct traceable classification evidence.
50%UnverifiedNarraSpace achieves an average accuracy of 80.88%±2.56% across 50 random seeds on the official 48-person test set of the ADReSS benchmark.
50%UnverifiedAfter averaging probabilities, NarraSpace achieves 81.25% accuracy and an AUC of 91.32% on the ADReSS test set.
50%UnverifiedNarraSpace's classification stage uses a frozen language model to compare text for each scene event, combined with transcript surprisal and coverage measurements, and finally predicts with a fixed SVM classifier.
50%UnverifiedNarraSpace retains source records, and every event score can be linked back to its source utterance, speech interval, and lexical match record.
50%UnverifiedNarraSpace's evaluation is based on the 48-person test set of the ADReSS benchmark, so the sample size is relatively limited, and the coverage of predefined scene events may also be constrained by the design of the image itself.
50%