This paper presents a proof of concept for realtime medical signal processing using audio as the transmission medium. Vital sign such ECG and PPG are used as an example, while the approach can be applied to any biomedical signal that changes slowly or moderately over time. The system architecture includes a transmitter and a receiver. The transmitter senses the medical signal, encodes it, and converts it into an audio waveform, which is transmitted via wire, air, or fluid. The receiver captures the sound, converts it back into an electrical and digital signal, decodes it, and processes it further. Low-cost components such as standard piezo speakers and microphones are used for signal conversion, while the processing algorithms run on affordable hardware or any mobile device. For airborne transmission, the system operates in the 16-22 kHz frequency range, which is typically inaudible or barely audible to humans. The system is capable of transmitting analog or digital signals, in single or multi-channel modes. The paper describes HeatlthTalk methodology on the system level with emphasis to software decoding on receiver side, implemented in open- source, platform-independent technologies such as JavaScript + CSS + HTML. The HealthTalk is low-cost, energy-efficient, easy to implement, eco friendly and robust against noise. It can operate without a persistent connection, making it suitable for "plugout" or offline environments. HealthTalk offers a potential alternative to wireless communication systems such as RF, Bluetooth (BT), and Bluetooth Low Energy (BLE) in appropriate applications.
The document discusses the development and evaluation of machine learning models for detecting credit card fraud in an imbalanced dataset, specifically using XGBoost and Random Cut Forest (RCF) algorithms. Credit card companies must identify fraudulent transactions to avoid charging customers for unauthorized purchases, using a dataset with 0.173% fraudulent examples out of 284,807 total examples. The initial solution used Amazon SageMaker and the RCF algorithm for unsupervised anomaly detection, which identified anomalies based on their characteristics and provided anomaly scores for test cases. The XGBoost supervised learning method with weighting schemes outperformed the RCF approach, with further enhancements possible using techniques like SMOTE and Hyperparameter Optimization (HPO). Improvements to the XGBoost algorithm involved transforming the binary problem into a multi-class problem using the Majority Oversampling Technique (MOTE), leading to better performance metrics. The evaluation metrics for the XGBoost MOTE model included balanced accuracy, Cohen's Kappa, F1 score, logistic loss functions, balanced logistic loss functions, and AUC, showing significant improvements over other models. The XGBoost MOTE model was compared with other models like XGBoost with built-in weighting, XGBoost with SMOTE, and RCF, showing superior performance in most metrics.
The deployment of deep learning models on edge devices is characterized by significant challenges due to limitations in computational power, memory and energy efficiency. In particular, quantization has evolved into a leading technique for reducing the size of the model and improving the speed of inference, while still mitigating the accuracy drop. This study analyzes the performance of quantization as an optimization method and explores the feasibility of training ML models directly on IoT devices. It focuses on the impact of quantization on two neural network models: LSTM (Long Short-Term Memory) and FNN (Feedforward Neural Network) using three schemes (tf, Float32, Int8). In a comparative analysis, throughput, memory consumption, latency and symmetric mean absolute percentage error (SMAPE) were used. The results show that Int8 quantization substantially reduces memory usage and improves latency and throughput, especially for the FNN model, which achieved 99,000 samples per second at a latency of just 0.01 ms. Despite its stable performance characteristics, the Int8 scheme provided benefits to the LSTM model. Deep learning models achieve better efficiency in real-time applications through Int8 quantization while preserving accuracy with minimal loss and gaining substantial performance benefits.
In this paper, we propose a hybrid system suitable for time-frequency (TF) signal analysis. This system applies the S-method for TF signal analysis and introduces an architecture that balances two traditional extremes. A single-clock-cycle (SCI) design delivers the fastest execution but demands prohibitively high hardware complexity, whereas a multiple-clock-cycle (MCI) design reduces hardware requirements at the cost of slower execution. The proposed solution blends the strengths of both approaches, trimming hardware complexity while keeping execution time short enough for practical use. Trade-offs and comparisons of the proposed design regarding the critical design performance and the competitive designs are also provided.
Information security helps in protecting the confidentiality, authenticity and integrity of user data. In today’s digital era, steganography within the field of information security has become an important data hiding technique, in this way addressing the need for secure data transmission. It enhances information security by providing an additional layer of confidentiality, allowing the sensitive data to be hidden within different media to protect it from detection or access. The type of media that steganography uses to protect the data are videos, audios and images, which is the media carrier that we are going to use. Meanwhile, image classification helps large image datasets to be processed efficiently with machine learning and deep learning models becoming essential in this area. We are going to evaluate the impact of steganography on the performance of machine learning models like Support Vector Machine, K-Nearest Neighbors, Logistic Regression and Random Forests, but also deep learning models like Residual Network with 50 layers and 18 layers, Vision Transformer and Visual Geometry Group with 16 layers.
