Mental health disorders affect a significant portion of the global population, making early detection and prevention an urgent challenge. Functional Magnetic Resonance Imaging (fMRI) offers a non-invasive means of mapping brain activity, while machine learning (ML) techniques provide powerful tools for analyzing its high-dimensional data. This paper presents a focused review of studies published between 2018 and 2023 that apply ML algorithms to fMRI data in the context of mental health. Unlike broader surveys, our approach selectively examines recent contributions, emphasizing the disorders investigated, the algorithms employed, and the reported performance metrics. We synthesize results across methods such as Support Vector Machines, Random Forests, and Neural Networks, highlighting both their advantages and limitations. By consolidating findings into comparative tables, we aim to provide a structured overview that clarifies current trends, methodological challenges, and promising directions. This work contributes a Computer Science–oriented perspective on the use of ML in fMRI analysis, identifying open issues related to generalizability, interpretability, and scalability that must be addressed in future research.
This study evaluates three pitch detection algorithms—YIN, MComb, and Schmitt Trigger—using one-second recordings of single musical notes and computer-generated sinusoidal signals. The use of short notes was motivated by applications such as automatic transcription, tuning systems, and real-time music analysis, which require algorithms to identify the fundamental frequency quickly and reliably. By isolating one-second single notes, we assessed pitch detection behavior under controlled conditions, without confounding factors such as overlapping notes, long-term dynamics, or tempo fluctuations. The tests included clean tones, signals with controlled distortions such as additional harmonic components or slight vibrato, background noise, and real saxophone recordings. Results showed that YIN achieved the highest accuracy under the tested synthetic and limited real-recording conditions, MComb showed strong performance on the evaluated real instrument recordings, and the Schmitt Trigger achieved high precision when noise levels were minimal. To address the limitations of individual methods, we implemented and tested a hybrid octave-aware weighted voting approach, which serves as a proof-of-concept combination strategy evaluated within this experimental setup.
This study examines the risk of QR code phishing — which we will be referring to as quishing from this point forward — in the context of political elections in Romania. Through a field experiment, we deployed posters featuring both political and social content, each framed in either a neutral or clickbait style, to evaluate how different social engineering techniques influence interaction outcomes. Our results indicate that politically framed clickbait posters generated substantially higher scan rates, while conversion patterns varied across conditions. These findings highlight the viability of QR codes as a powerful attack vector in politically charged environments, and the need for awareness campaigns, as subjects were not only interested in scanning them, but also in submitting their email addresses.
We investigate the problem of breast cancer screening optimization, using various techniques applicable in domains where the data format is not defined in advance. The aim is to minimize the cost related to the screening of patients while maximizing the beneficial effect of the process regarding some key breast cancer indices. Our model can be easily adjusted to other similar network optimization tasks where a goal function has to be minimized across a geographical surface. We present the problem’s key similarities to the Travelling Salesman Problem and underline the fact why we choose a deterministic algorithm compared to a Simulated Annealing-based solution. Furthermore, we present the usefulness of the Elastic Stack regarding this application and offer a concrete solution to the problem defined by our generated dataset, respecting the European data distributions in this domain.
The paper addresses the topic of detecting money laundering operations in transaction data represented as graph data-structures. We propose the integration of autoencoder components in Graph Neural Networks (GNN) architectures, in order to incorporate a reconstruction step in the traditional edge classification problem and enhance model quality based upon the usage of reconstruction errors. We show that enhancing GNNs with autoencoder components improves the predictive performance of money laundering detection, on data represented as homogeneous graphs. Additionally, the Shapley value is computed in order to gain further insight into the most important features from distinguishing normal and fraudulent activities.
