AMS-HD: Acute Mountain Sickness Detection with Hyperdimensional Computing

Abu Kaisar Mohammad Masum^✢, Reeti Pradhananga^✢, Jonas I. Schmidt^✢,

Mehran Moghadam^✥, M. Hassan Najafi^✥, Bige Unluturk^✝,

Ulkuhan Guler^✠, Sercan Aygun^✢

✢School of Computing and Informatics, University of Louisiana at Lafayette, Lafayette, LA, USA

✥Electrical, Computer, and Systems Engineering Department, Case Western Reserve University, Cleveland, OH, USA

✝Electrical and Computer Engineering & Biomedical Engineering, Michigan State University, East Lansing, MI, USA

✠Electrical & Computer Engineering, Worcester Polytechnic Institute, Worcester, MA, USA

ID:1359

Agenda

• Abstract
• Introduction
• Background
• Proposed AMS-HD
• Experimental results
• Conclusion

​

​

​

Abstract

• AMS (Acute Mountain Sickness) is potentially life-threatening
• Traditional machine learning for AMS detection have limitations
• HDC(Hyperdimensional Computing) has not yet been applied to AMS detection with limited biomedical data
• HDC offers high accuracy with low hardware complexity, ideal for resource-constrained devices like wearables
• A novel framework, AMS-HD, is proposed using:
• Custom feature engineering
• Quasi-random hypervector encoding
• AMS-HD enables real-time AMS detection and can be integrated into wearable devices for continuous health monitoring

​

​

Abu Kaisar Mohammad Masum et al. (2025)

Introduction

• AMS occurs at altitudes above 2,500 meters due to low oxygen
• Common symptoms: headache, nausea, tiredness, vomiting—can appear 6–12 hours after ascent
• Early detection is critical to prevent severe complications
• Wearable devices offer a lightweight, real-time solution
• Traditional ML methods are computationally intensive and less suited for resource-constrained environments
• HDC is introduced as a fast, efficient, and reliable alternative

Abu Kaisar Mohammad Masum et al. (2025)

Introduction

• HDC mimics brain-like functionality using high-dimensional vectors for robust, noise-tolerant processing.

​

• HDC uses binary or bipolar hypervectors with operations like XOR, Add, Shift, and Permute

Abu Kaisar Mohammad Masum et al. (2025)

Background

• Conventional ML in AMS Prediction
• ML models predict AMS severity using:�SpO₂ (peripheral oxygen saturation), heart rate, etc.
• Genetic data-based models exist but are complex.
• Common machine learning methods:�logistic regression, SVM, KNN, bagged trees

​

• Limitations: High computational requirements, unsuitable for low-power or real-time applications

Abu Kaisar Mohammad Masum et al. (2025)

Background

• Hyperdimensional Computing (HDC)
• HDC represents data as high-dimensional vectors (HVs) with thousands of binary/bipolar values
• Core operations: Binding (XOR/multiplication), Bundling (summation), Permutation (reordering)
• HVs are nearly orthogonal and robust to noise/errors
• Pseudo-random HVs may reduce orthogonality; quasi-random (Sobol) sequences improve accuracy
• HDC supports efficient, low-memory, low-power real-time classification—ideal for wearables

Abu Kaisar Mohammad Masum et al. (2025)

Background

General HDC Framework

Abu Kaisar Mohammad Masum et al. (2025)

Proposed AMS-HD

• Dataset Overview
• Physiological responses (e.g., SpO₂, HR) at high altitude
• Selected features: SpO₂ (%), HR (bpm)
• AMS score range: 1 to 12
• Binary classification: AMS (score ≥ 3) vs. NO AMS (score < 3)
• Multiclass classification: Mild (3–5), Moderate (6–9), Severe (10–12)
• Feature Engineering
• Positional Encoding & Data Projection
• Hypervector Generation
• HD Classifier

​

Abu Kaisar Mohammad Masum et al. (2025)

Proposed AMS-HD

Feature Engineering

Positional Encoding & Data Projection

Features Selection

Abu Kaisar Mohammad Masum et al. (2025)

Proposed AMS-HD

HD Classifier

Hypervector (HV) Generation

Abu Kaisar Mohammad Masum et al. (2025)

Experimental results

• Binary Classification
• Logistic Regression (LR) and Support Vector Machine (SVM) 73% test accuracy.
• Random Forest (RF) 53% test accuracy
• HDC-Random and HDC-Quasirandom 67% test accuracy

Abu Kaisar Mohammad Masum et al. (2025)

Experimental results

• Multi Classification
• RF and SVM with 73% accuracy.
• LR with 55% accuracy
• HDC-Random achieved 64%, and HDC-Quasirandom achieved 73% accuracy

Abu Kaisar Mohammad Masum et al. (2025)

Experimental results

• Runtime. HDC exhibits the shortest training time, with the quasi-random method achieving the lowest inference time

Abu Kaisar Mohammad Masum et al. (2025)

(a)Binary

(b) Multiclass

Experimental results

• HDC-Q and HDC-R used less memory
• HDC-Q and HDC-R use less power
• HDC-Q achieves less energy

Abu Kaisar Mohammad Masum et al. (2025)

than Conventional models

Experimental results

• HDC model performance on both recall and F1-score

Abu Kaisar Mohammad Masum et al. (2025)

Conclusion

• HDC provides a powerful solution for machine learning tasks.

​

• AMS-HD is an HDC-based framework designed for�1 real-time�2 accurate�3 resource-efficient detection of Acute Mountain Sickness (AMS)

​

• AMS-HD is highly suitable for real-time health monitoring

​

• AMS-HD can be integrated with wearable and IoT devices

Abu Kaisar Mohammad Masum et al. (2025)

Abu Kaisar Mohammad Masum et al. (2025)

Email to: Abu Kaisar Mohammad Masum c00591145@louisiana.edu