Simulation studies were performed to assess the performances in realistic estimation problems of B-splines (BS) and restricted cubic splines (RCS) within flexible models for survival data. Several theoretical distributions, designed to match the estimates from clinical cancer studies, were adopted for data generation, including a proportional hazards (PH) model and a non-PH one. The simulation plan included sample sizes equal to 500 and 2000, with 10 ^2 ), and the ability to identify the correct number of peaks of the target hazard, along with the ability to detect peak times and peak heights within clinically relevant ranges. We found reduced performances in the pattern identification task, even for the simplest distributional models, for which estimation error was optimal. Overall, BS slightly outperformed RCS in pattern identification, while estimation error was equivalent. These results will be compared to further simulations concerning a more extensive set of smoothing methods.
Furthermore, accelerating the process of interpretation of forensics DNA analysis can be crucial for faster and more precise identification, especially in cases where time and accuracy are critical, and improving forensic analysis and decision-making.
This study aims to evaluate haematological parameters of patients who have contracted the COVID-19 infection. In particular, we considered patients who were already hospitalised at the time of nasopharyngeal swab (NF) testing, and patients at the Emergency Department and/or who required hospitalisation following a positive result from the NF swab. The collected data are defined as longitudinal data (i.e. constituted by measurements accumulated sequentially over time), mainly characterised by observation times that are irregular, different for each patient, and distributed non-uniformly across the observation interval. In light of these considerations, we exploit CONNECTOR, a data-driven framework designed for longitudinal data, which returns a grouping of the curves of the haemochromocytometric parameters, based on a functional clustering algorithm. These clusters are analysed in terms of disease outcome and survival, finding a good correlation to mortality rates. Finally, the CONNECTOR clusters are exploited to stratify the patients based on profiles of the haemochromocytometric parameters evolution over time, showing that comorbidities appear to have an impact on mortality independently of the outcome of the monitoring carried out through the laboratory tests considered.
In medical imaging, several deep learning models, such as vision transformers (ViT), have shown improved capability in recognizing image patterns efficiently by enhancing model efficiency in parameter optimization and sample effectiveness. Our research introduces a fusion approach that leverages multi-backbone pre-trained models (ResNet, EfficientNet, VGG) as feature extractors and the Caenorhabditis elegans' pyramid connectome ViT in the tail. In most cases, the proposed model demonstrates superior performance on MEDMNIST2D V2.0 classification challenges. Moreover, our approach maintains comparable performance while utilizing fewer training parameters than conventional state-of-the-art models. This indicates a significant step toward more efficient deep-learning architectures in medical diagnostics.
Feature (variable) selection methods are used to detect the most important features (variables) within high-dimensional data. Tree-based models, such as Random Forests, are often exploited for this purpose, as they provide a built-in mechanism to quantify feature importance. However, the stochastic sampling strategies used in these models can lead to unstable feature importance rankings, particularly when the number of trees is low. In our study, we investigate the extent to which these unstable feature rankings can be consolidated through rank aggregation and consensus signal techniques. We propose to compute consensus values from multiple feature selection runs, where each run generates a ranked list of features. We have evaluated our approach while varying a spectrum of hyperparameters such as the number of trees and the number of features available for splitting a node. Our results suggest that consensus ranks provide a more accurate and robust selection of features compared to single-run feature selection procedures. The proposed approach is especially relevant for biomarker discovery, as it can improve the accuracy and reliability of feature selection, leading to the identification of the most informative and relevant biomarkers associated with a particular disease or condition. By consolidating the results from multiple feature selection runs, our approach may help to overcome problems associated with noisy or complex data because it can provide more robust and accurate estimates of the feature importance rankings.
