nterpretable Alzheimer’s Disease Detection from CHAT Transcripts Using Criterion-Guided LLM Prompting and In-Context Learning

The human voice carries more than just words; it holds a map of the mind. For decades, researchers have listened to the speech of people with Alzheimer's disease, searching for subtle shifts in how they choose words, how they pause, and how they weave their stories together. These changes often appear long before memory loss becomes obvious, offering a potential window for early detection. Traditionally, finding these patterns required building complex computer programs that needed to be taught with thousands of labeled examples, a process that was slow, expensive, and often difficult to explain to a doctor. The field has recently seen a shift toward using large language models, powerful artificial intelligence systems trained on vast amounts of text. Unlike older programs, these models can understand context and reason through problems without needing to be retrained on every new task, provided they are guided correctly. The challenge has been figuring out how to make these general-purpose tools listen to the specific, messy details of clinical speech and translate them into a reliable diagnosis.

A team of researchers has developed a new way to use these AI tools to detect Alzheimer's disease from spoken transcripts, achieving high accuracy without teaching the AI anything new. Instead of forcing the computer to learn from scratch, they designed a step-by-step guide that mimics how a human expert might analyze a patient's speech. The process begins with the raw text of a conversation, which often contains strange symbols used by linguists to mark pauses, repetitions, and unclear sounds. The researchers first converted these symbols into plain English sentences, turning technical notes like "short pause" or "repetition" into readable text that the AI could understand naturally. This step was crucial, as it transformed the data from a coded format into a story the machine could actually read.

Once the text was prepared, the system asked the AI to act as a clinical observer, breaking down the analysis into three specific questions. First, it looked at the vocabulary: was the speaker using precise words to name objects, or were they relying on vague terms like "thing" or "stuff"? Second, it examined the flow of speech for signs of struggle, such as frequent repetitions, self-corrections, or long silences. Third, it checked the overall story: did the speaker stay on topic and connect their ideas logically, or did the narrative drift and fall apart? By answering these three questions, the AI built a reasoned argument for its final decision, rather than just guessing a label. To help the AI make the right call, the researchers also provided it with a few examples of past conversations that covered a wide range of difficulties, from very clear cases to confusing ones, ensuring the AI saw the full picture of what the disease looks like in speech.

The results of this approach were striking. When tested on a standard set of 48 recorded conversations, the system correctly identified the condition in 85.4 percent of the cases, outperforming previous methods that relied on training the computer with labeled data. More importantly, the system did not just give a yes-or-no answer; it provided a clear explanation for its decision, citing the specific hesitations or vague words that led to its conclusion. This transparency is vital for medical use, as it allows doctors to see the reasoning behind the diagnosis. However, the study also revealed a significant limitation: the system's success depended entirely on the quality of the text. When the researchers tried using transcripts generated automatically by speech-to-text software, the accuracy dropped sharply. The automated systems failed to capture the very pauses and repetitions that the AI needed to spot the disease, proving that for now, human-transcribed notes remain essential for this method to work.

The researchers found that the way they guided the AI mattered more than the size of the AI itself. Simply asking the computer to guess the diagnosis in one step resulted in much lower accuracy. It was the step-by-step reasoning, combined with the carefully chosen examples, that unlocked the system's ability to detect the subtle signs of cognitive decline. The study suggests that with the right prompts and a clear understanding of clinical criteria, artificial intelligence can be a powerful, transparent tool for early detection, even without the need for complex training. While the current method requires human-transcribed speech, the success of this approach points toward a future where AI can assist clinicians in identifying Alzheimer's disease earlier and more reliably, provided the input data is precise enough to carry the weight of the diagnosis.