Scenario-Based Generative AI for Early Detection and Management of Anxiety Attacks in Dementia Care

Abstract

Purpose: Dementia is linked with various neuropsychiatric challenges, including anxiety attacks. Managing these attacks in individuals with dementia presents unique complexities due to their potential difficulty in communicating emotions or following instructions(Kwak et al., 2017). Expert caregivers possess skills in detecting the early signs of an anxiety episode, removing triggers promptly, accurately predicting the attack's progression, and intervening to comfort the patient. Such expertise is vital in managing the mental health of dementia patients, alleviating the occurrence of anxiety episodes, and controlling their damages. Technological support can enhance early detection of anxiety attacks in individuals with dementia(Tsai et al., 2022). It can also identify potential scenarios and recommend appropriate responses or interventions based on these scenarios. This involves using AI models to analyze patterns and signs indicating the onset of an anxiety attack. This approach aims to prepare caregivers with the right strategies to prevent the attack or lessen its impact, thereby improving the care and support to individuals with dementia. Method: We have developed the outline of a generative AI model that maps out the sequence of events leading from various triggers to the onset of anxiety attacks in patients with dementia (Figure 1). It creates scenarios to forecast the occurrence of anxiety attacks prompted by these triggers. The scenarios are created as time-series of tokens, each representing a specific event or condition. These token sequences can be translated into simple sentences, making them easy to understand. Based on these events, the system provides suggestions for intervention. In its training phase, the model captures time series data of stimuli and anxiety attack incidents. It specifically monitors a combination of daily routine timings and significant sensory stimuli, like loud noises or reports of pain. It also tracks or receives reports on environmental changes or disruptions in daily routines. The model then learns to associate these stimuli with recorded incidents of anxiety attacks, including the attack's delay, duration, and both physical symptoms (like increased heartbeat, sweating, trembling, shortness of breath, reported pains, and nausea) and behavioral responses (such as avoiding specific locations). During the training phase, the model identifies the recurrence patterns of anxiety attack incidents over time in presence of stimuli. In the application phase, the model begins by processing real-time data on the occurrence of stimuli. It then generates a sequence predicting possible upcoming anxiety attack events. Subsequently, this current stream of stimuli data is combined with previously predicted sequences of anxiety events and reintroduced into the model for enhanced prediction follow-up accuracy. These predictions are then communicated to caregivers, who provide feedback based on the actual occurrence of the events. Should the predictions deviate from real events beyond a predetermined threshold, the model resets to a default state, assuming a 'no attack' scenario, and starts anew. A series of early intervention guidelines is established, synchronized with the progression of events. When the model enters an attack-alarm state, it suggests specific interventions. These recommended behavioral interventions are tailored to the most likely scenarios as forecasted by the model, aiming to find the most direct and effective path to restore the individual's anxiety level to its normal baseline. Results and Discussions: We outlined a generative AI model that identifies the timing patterns and associations between various stimuli occurring in the environment or internally to the patient and the onset of anxiety attacks. The model has the ability to simulate and predictively regenerate the sequence of anxiety attack events likely triggered by these stimuli. It is highly adaptable, enabling fine-tuning for individual patients to accommodate personal differences, thus aligning with precision health principles.

References

Kwak, Y. T., Yang, Y., & Koo, M.-S. (2017). Anxiety in Dementia. Dementia and Neurocognitive Disorders, 16(2), 33–39. https://doi.org/10.12779/dnd.2017.16.2.33

Tsai, C.-H., Chen, P.-C., Liu, D.-S., Kuo, Y.-Y., Hsieh, T.-T., Chiang, D.-L., Lai, F., & Wu, C.-T. (2022). Panic Attack Prediction Using Wearable Devices and Machine Learning: Development and Cohort Study. JMIR Medical Informatics, 10(2), e33063. https://doi.org/10.2196/33063