TCP-E-SNN: Temporal Confidence Propagation with Engram Memory for Calibrated Fall Detection Across Patient Demographics

EasyChair Preprint 16029

Abstract

Wearable fall detection systems face a critical deployment challenge: models trained on younger patient populations produce miscalibrated confidence scores when deployed for elderly patients,making their uncertainty estimates unreliable for clinical decision- making. We present TCP-E-SNN (Temporal Confidence Propagation with Engram Memory Spiking Neural Network), a neuromorphic architecture that addresses demographic distribution shift through two biologically grounded mechanisms. First, a Temporal Confidence Propagation (TCP) module derives per-sample confidence directly from spike-timing dynamics, inter-spike interval(ISI) variance, and first-spike latency. This encodes uncertainty as a signal independent of the output layer’s softmax distribution. Second, an Engram Memory module maintains =16 Hebbian-updated prototype patterns, injecting priors into the LSTM initial hidden state at inference time, enabling the model to distinguish familiar from genuinely ambiguous spike patterns. We validated on SisFall(153,705 windows, 38 subjects) under two conditions: a standard 80/20 split and a cross-age distribution shift (trained on young adults SA01–SA23, tested on elderly SE01–SE15). Under distribution shift, TCP-E-SNN achieved a 62.9% reduction in Expected Calibration Error (ECE 0.0294 vs. 0.0792 for a temperature-scaled Rate-SNN baseline), while also outperforming the baseline on F1 by +2.74 pp (0.7487 vs. 0.7213). On the standard in-distribution evaluation, TCP-E-SNN achieves a 17.6% ECE reduction (ECE 0.0114vs. 0.0138) with a +1.82 pp F1 advantage (0.8125 vs. 0.7943). These results demonstrate that biologically-inspired temporal confidence signals, augmented with associative engram memory, provide substantially more reliable uncertainty estimates under demographic distribution shift than softmax-based calibration alone. This results in direct implications for trustworthy AI in elderly care.

Keyphrases: Healthcare AI, Hebbian plasticity, Inter-Spike Interval, Spiking Neural Networks, calibration, distribution shift, early exit inference, engram memory, fall detection, neuromorphic computing, quantification, uncertainty

@booklet{EasyChair:16029,
  author    = {Blessing Effiong and Veric Tan},
  title     = {TCP-E-SNN: Temporal Confidence Propagation with Engram Memory for Calibrated Fall Detection Across Patient Demographics},
  howpublished = {EasyChair Preprint 16029},
  year      = {EasyChair, 2026}}