Appworks #33

Traditionally, dementia has often been conceptualized as a single disease entity requiring a single therapeutic solution. Clinical experience, however, suggests a more complex reality. Cognitive decline in older adults is rarely a smooth, uniform process. Instead, it often emerges from multiple interacting vulnerability nodes, including infection, hospitalization, delirium, traumatic brain injury, sleep disruption, medication changes, vascular events, affective symptoms, systemic inflammation, and caregiver-contextual stressors. These nodes may not act continuously or linearly; rather, they may become clinically visible at specific moments when a patient abruptly fails to return to a previous baseline.

Similarly, dementia progression is often imagined as a gradual downward slope. Yet in real-world care, many families describe a different pattern: a stepwise or cliff-like deterioration after a triggering event. A caregiver may report that an older adult “recovered from pneumonia, but never came back,” “became slower after a fall,” or “was not the same person after hospitalization.” Such descriptions may not fit neatly into standardized diagnostic language, but they often mark clinically meaningful transitions in memory, attention, executive function, behavior, sleep, and daily function. These caregiver observations may therefore provide early signals of rapid deterioration at precisely the moments when intervention, follow-up, or reassessment may still alter the subsequent trajectory.

The aim of NarrativeBio is not to position artificial intelligence as an autonomous diagnostic system for dementia. Rather, it is to support the timely recognition of clinically meaningful deterioration points. If rapid decline can be captured at the moment it occurs, caregiver narratives may help identify the relevant event history, temporal pattern, functional change, and candidate neurocognitive risk axes. In this framework, the question is not simply whether an individual “has dementia,” but how, when, and through which vulnerability nodes the person may be deteriorating. By transforming unstructured caregiver narratives into explainable, clinician-verifiable neurocognitive risk profiles, AI may help clinicians and caregivers better recognize where the patient is falling, what domains are changing, and whether the trajectory of decline can still be modified.

To operationalize this vision, this Technical Report proposes a five-pillar framework for human-in-the-loop explainable AI in caregiver narrative structuring. The five pillars are: (1) caregiver narratives as ecologically rich clinical signals; (2) theory-guided neurocognitive mapping; (3) explainable AI with evidence tracing; (4) human-in-the-loop expert verification; and (5) validation and accountable scalability. The framework is designed to convert caregiver-described changes into structured, auditable, and clinically cautious outputs without presenting AI as a substitute for clinical diagnosis or expert judgment.