Background: Breast cancer is one of the major causes of mortality globally. Therefore, different Machine Learning (ML) techniques were deployed for computing survival and diagnosis. Survival analysis methods are used to compute survival probability and the most important factors affecting that probability. Most survival analysis methods are used to deal with clinical features (up to hundreds), hence applying survival analysis methods like cox regression on RNAseq microarray data with many features (up to thousands) is considered a major challenge.Methods: In this paper, a novel approach applying autoencoder to reduce the number of features is proposed. Our approach works on features reconstruction, and removal of noise within the data and features with zero variance across the samples, which facilitates extraction of features with the highest variances (across the samples) that most influence the survival probabilities. Then, it estimates the survival probability for each patient by applying random survival forests and cox regression. Applying the autoencoder on thousands of features takes a long time, thus our model is applied to the Graphical Processing Unit (GPU) in order to speed up the process. Finally, the model is evaluated and compared with the existing models on three different datasets in terms of run time, concordance index, and calibration curve, and the most related genes to survival are discovered. Finally, the biological pathways and GO molecular functions are analyzed for these significant genes.Results: We fine-tuned our autoencoder model on RNA-seq data of three datasets to train the weights in our survival prediction model, then using different samples in each dataset for testing the model. The results show that the proposed AutoCox and AutoRandom algorithms based on our feature selection autoencoder approach have better concordance index results comparing the most recent deep learning approaches when applied to each dataset. Each gene resulting from our autoencoder model weight is computed. The weights show the degree of effect for each gene upon the survival probability. For instance, four of the most survival-related experimentally validated genes are on the top of our discovered genes weights list, including PTPRG, MYST1, BG683264, and AK094562 for the breast cancer gene expression dataset. Our approach improves the survival analysis in terms of speeding up the process, enhancing the prediction accuracy, and reducing the error rate in the estimated survival probability.

Microarray technology is one of the most important recent breakthrough in experimental molecular biology. This novel technology for thousands of genes concurrently allows the supervising of expression levels in cells and has been increasingly used in cancer research to understand more of the molecular variations among tumors so that a more reliable classification becomes attainable. Machine learning techniques are loosely used to create substantial and precise classification models. In this paper, a function called Feature Reduction Classification Optimization (FeRCO) is proposed. FeRCO function uses machine learning techniques applied upon RNAseq microarray data for predicting whether the patient is diseased or not. The main purpose of FeRCO function is to define the minimum number of features using the most fitting reduction method along with classification technique that give the highest classification accuracy. These techniques include Support Vector Machine (SVM) both linear and kernel, Decision Trees, Random Forest, K-Nearest-Neighbors (KNN) and Naïve Bayes (NB). Principle Component Analysis (PCA) both linear and kernel, Linear Discriminant Analysis (LDA) and Factor Analysis (FA) along with different machine learning techniques were used to find a low-dimensional subspace with better discriminative features for better classification. The major outcomes of this research can be considered as a roadmap for interesting researchers in that field to be able to choose the most suitable machine learning algorithm whatever classification or reduction.

The proliferation of Internet of Things (IoT) devices challenges deep learning (DL) deployment due to their limited computational power, while cloud offloading introduces high latency and network strain. Fog computing provides a viable middle ground. We present a resource-aware framework that intelligently partitions DL tasks between fog nodes and the cloud using a novel Binary Search-Inspired Recursive (BSIR) optimization algorithm for rapid, low-overhead decision-making. This is enhanced by a novel module that fine-tunes deployment by analyzing memory at a per-layer level. For true adaptability, a Retrieval-Augmented Generation (RAG) technique consults a knowledge base to dynamically select the best optimization strategy. Our experiments demonstrate dramatic improvements over established metaheuristics. The complete framework boosts memory utilization in fog environments to a remarkable 99%, a substantial leap from the 85.25% achieved by standard algorithms like Genetic Algorithms (GA), Simulated Annealing (SA), and Particle Swarm Optimization (PSO). The enhancement module alone improves these traditional methods by over 13% without added computational cost. Our system consistently operates with a CPU footprint under 3% and makes decisions in fractions of a second, significantly outperforming recent methods in speed and resource efficiency. In contrast, recent DL methods may use 51% CPU and take over 90 s for the same task. This framework effectively reduces cloud dependency, offering a scalable solution for DL in the IoT landscape.