AMS-HD: Hyperdimensional Computing for Real-Time and Energy-Efficient Acute Mountain Sickness Detection

Technical Summary: AMS-HD: Hyperdimensional Computing for Real-Time and Energy-Efficient Acute Mountain Sickness Detection

Problem Statement
Acute Mountain Sickness (AMS) is a prevalent condition affecting individuals ascending above 2,500 meters, with symptoms ranging from headache and nausea to life-threatening cerebral or pulmonary edema. While continuous monitoring of physiological signals via wearables offers a promising avenue for early detection, conventional Machine Learning (ML) methods (e.g., Support Vector Machines, Multi-Layer Perceptrons) often fail to meet the strict real-time and energy-efficiency requirements of resource-constrained wearable and embedded devices. These traditional models typically demand significant computational resources, high memory footprints, and substantial power consumption, limiting their suitability for "always-on" health monitoring in remote, high-altitude environments.

Methodology: The AMS-HD Framework
The authors propose AMS-HD, the first Hyperdimensional Computing (HDC) framework specifically designed for real-time AMS detection. HDC is a brain-inspired paradigm that represents data as high-dimensional hypervectors (HV), enabling robust, noise-tolerant computation through simple vector operations (binding, bundling, permutation) rather than complex arithmetic.

The framework is unique in its dual-layer design, spanning from high-level software to low-level hardware:

1. High-Level Bipolar Computing (-1/+1): Designed for mobile platforms (smartphones and smartwatches), this layer utilizes bipolar vectors for efficient inference on general-purpose processors.
2. Low-Level Binary Computing (0/1): Optimized for Field-Programmable Gate Arrays (FPGA) and Application-Specific Integrated Circuits (ASIC), this layer uses binary logic to minimize hardware resource usage and power consumption.

Key Technical Components:

• Feature Selection: The system utilizes Mutual Information (MI) to identify the most predictive physiological features. Analysis confirmed that Oxygen Saturation (SpO2) and Heart Rate (HR) are the dominant predictors, while blood-derived variables (e.g., Hematocrit, Carbon Monoxide) were excluded from the real-time pipeline as they are not wearable-accessible.
• Encoding Strategies: The framework employs various encoding techniques to generate hypervectors, including pseudo-random, quasi-random (Sobol sequences), and Hadamard-based encodings. A novel thermometer (unary) encoding is implemented in the hardware design to map continuous scalar features (like SpO2 and HR) to binary vectors, preserving numerical proximity and correlation.
• Positional Encoding: To handle temporal and event-based context (e.g., baseline vs. night vs. overnight stages), the system integrates positional encoding, allowing the model to distinguish between different stages of altitude exposure without requiring dense continuous time-series streams in the current dataset evaluation.
• Classification: The model performs both binary classification (AMS vs. No AMS) and multiclass classification (severity levels: Mild, Moderate, Severe). Inference is conducted via similarity search (e.g., cosine similarity or Hamming distance) between a query hypervector and stored class hypervectors.

Key Contributions

1. First HDC Framework for AMS: Introduction of AMS-HD as the first complete HDC-based solution for altitude sickness detection, bridging the gap between wearable inference and low-level hardware deployment.
2. Dual-Platform Architecture: A unified framework supporting high-level bipolar computing for mobile devices and low-level binary computing for hardware accelerators (FPGA/ASIC), enabling deployment across diverse resource-constrained environments.
3. Optimized Feature Engineering: Implementation of MI-based feature selection and specialized encoding strategies (including quasi-random Sobol sequences and unary encoding) to enhance detection accuracy and energy efficiency.
4. Comprehensive Validation: Extensive evaluation across three distinct platforms: an FPGA prototype (PYNQ-Z2), an ARM-based embedded processor, and a smartwatch–smartphone system (Samsung Galaxy Watch4/S21 Ultra).

Results
The evaluation demonstrates that AMS-HD achieves competitive or superior performance compared to conventional ML baselines (SVM and MLP) while drastically reducing resource consumption:

• Accuracy: On the primary dataset, AMS-HD achieved up to 91% accuracy and a 90% F1-score in binary classification (using Sobol + Unary encoding with D=256). It also achieved 85% accuracy on external AMS-related datasets without retraining, demonstrating strong generalization.
• FPGA Performance: On the PYNQ-Z2 FPGA, AMS-HD reduced Look-Up Table (LUT) usage by 7.3× and Flip-Flop (FF) usage by 5.8× compared to an MLP. It consumed 3.9× less power than the MLP baseline (1.6 W vs. 6.2 W) and required only 16 clock cycles for inference.
• Mobile Performance: On mobile platforms, AMS-HD required only 1% battery drop per session, utilized 60 Bytes of memory, and achieved a 2.50 ms inference time. This represents approximately 2× lower energy consumption than SVM and >3× lower than MLP.
• Multiclass Capability: The framework successfully extended to multiclass severity classification, achieving up to 76% accuracy, enabling finer-grained monitoring of AMS progression.

Significance
The paper claims that AMS-HD represents a transformative step for resource-constrained biomedical applications. By leveraging the inherent efficiency of hyperdimensional computing, the framework provides a scalable, hardware-aware alternative to traditional ML. It enables always-on, real-time monitoring of AMS without reliance on cloud connectivity or heavy computational resources. This work establishes a foundation for proactive management of high-altitude illness, offering a practical pathway for deploying continuous health monitoring systems in remote and challenging environments where power and processing capabilities are limited. The authors position this as a complete design flow from algorithmic concept to hardware implementation, addressing a critical gap in current wearable health technologies.