Personalised modelling with spiking neural networks integrating temporal and static information

Area of Science:

• * Computational neuroscience and machine learning.
• * Development of intelligent systems for personalized health analytics.

Background:

• * Existing prognostic/diagnostic systems often struggle with integrating diverse data types like static and spatiotemporal features.
• * Need for personalized models that can interpret complex interactions within individual health data.

Purpose of the Study:

• * To propose a novel personalized prognostic/diagnostic system integrating static and dynamic/spatiotemporal data.
• * To enhance classification, prediction, and pattern recognition accuracy for individual health assessments.
• * To improve the interpretability of predictive models for end-users.

Main Methods:

• * Development of a d2WKNN clustering method for optimal selection of neighboring samples based on integrated static and temporal individual data.
• * Training of a Personalised Spiking Neural Network (PSNN) using selected relevant samples to learn spatiotemporal patterns from streaming data.
• * Individualized optimization of system hyper-parameters, including PSNN and d2WKNN clustering parameters.

Main Results:

• * Achieved higher classification/prediction accuracy (80% to 93%) compared to global modeling and conventional methods.
• * Demonstrated the ability of PSNN models to capture complex space and time association patterns.
• * Successfully applied the system to neuroimaging data for treatment response classification and environmental data for stroke risk prediction.

Conclusions:

• * The proposed PSNN model offers a powerful analytical tool for personalized health applications.
• * The system enhances understanding of feature interactions, leading to better decision-making in healthcare.
• * Individualized profiling supports a deeper comprehension of health conditions and treatment responses.