Discharge summaries are free-text notes that outline the hospital stay of a patient, including diagnoses, treatments, and follow-up care recommendations. Through the use of International Classification of Diseases (ICD) coding, they play a vital role in communicating clinical data exchange and reimbursement. However, manually assigning ICD codes to discharge summaries can be labor-intensive and prone to errors due to the unstructured narrative format, variations in terminology, potential inaccuracies in documentation, and the limitations of the ICD coding system in capturing the complexity of patient conditions. This paper aims to improve the results of state-of-the-art automated medical coding machine learning models with two preprocessing techniques. The first technique involves data management to maximize the amount of usable data by mapping ICD-9 codes in the MIMIC-IV dataset to their corresponding ICD-101. With a similar goal, the second technique leverages large language models (LLMs) for the generation of multiple synthetic discharge summaries corresponding to ICD codes with few appearances in the dataset. The results of this extensive preprocessing are evaluated with the use of several well-established and state-of-the-art deep learning models, which present significant improvement.

In this work, we investigate the relation between weather and atmospheric conditions, and the biomarkers Platelet-Lymphocyte Ratio (PLR) and Neutrophil-Lymphocyte Ratio (NLR). In more detail, we are interested in identifying how significant weather and atmospheric changes can affect people’s health over time. For this purpose we introduce a new dataset that combines data from a recent population study along with historical meteorological measurements. We initially conduct timeseries exploratory analysis and then focus on changepoint detection and analysis of multivariate time series data in an effort to identify major incidents and their consequences. Interestingly, the results showed that the NLR and PLR biomarkers are significantly related to the major changes in weather and atmospheric conditions.

This paper introduces a comprehensive framework for clustering analysis, centered on a novel incremental silhouette score calculation designed specifically for hierarchical clustering. This innovative method significantly reduces the computational complexity of silhouette evaluation, transforming the process from O(K N) to effectively O(N) for K hierarchical configurations (demonstrated by an over 100-fold speedup in our tests) making it feasible for large-scale datasets and enabling efficient cluster number estimation within hierarchical clustering scenarios. Building on this, we revisit and enhance the Principal Direction Divisive Partitioning (IPDDP) algorithm, proposing principal component analysis-maximum margin divisive clustering (PCA-MMDC), which utilizes multiple principal components for more accurate data partitioning, and PCA-MMDC-sc, which incorporates a scatter-based cluster selection for improved balance. These are integrated into a hybrid clustering strategy that combines the strengths of incremental silhouette calculation and the enhanced algorithms, allowing for robust cluster identification and effective management of noise and outliers. Experimental results on synthetic and real-world datasets demonstrate notable improvements in clustering accuracy (achieving an average Adjusted Rand Index (ARI) increase of over 10 percentage points on custom noisy synthetic datasets compared to K-Means) and computational efficiency. While the choice of principal components in PCA-MMDC presents a parameter, the overall framework offers a scalable and robust solution for complex clustering tasks, with future work aimed at adaptive parameter selection and extending incremental calculations to other validation metrics.

This research looks at the use of long-short-term memory (LSTM) networks to predict psychosis, in patients within the schizophrenia spectrum, based on Heart Rate Variability (HRV) data acquired from wearable devices. Our primary objective is to test whether the personalized relapse prediction remains accurate when eliminating the artifact-removal preprocessing. We first analyzed 7 patients sleep HRV recordings (7–113 days each), and then validated the methodology on a separate 30-patient psychosis cohort from another clinical setting. In this framework, HRV characteristics are computed directly from the unprocessed time series for each patient, without artifact correction, at any stage prior to feature extraction. HRV features are then, organized into sequential inputs, where the model uses the first n−1 steps to predict the nth step. This structure allows the model to learn from temporal relationships and individual physiological trends in HRV. The sequence length used by the LSTM is optimized for each patient, allowing the model to account for individual physiological patterns. Through this, on the 7 patient cohort, the LSTM model reaches a mean F1 score of 0.9817, marking its strength across diverse patient profiles. The method provides predictions for each individual by learning from their own HRV history. Results using both traditional and state-of-the-art noise-removal techniques, like wavelet and GAN-based denoising, showed that omitting these data cleaning steps did not reduce, and in some cases even improved, prediction accuracy. These findings indicate that, for psychosis prediction based on wearable HRV data, additional data cleaning may not be necessary.

Abstract The utilization of large language models (LLMs) in research is becoming increasingly prevalent, as they offer advanced capabilities in processing and generating human-like text. However, this advancement comes with a significant trade-off in terms of time and computational costs. In this paper, we demonstrate that analyzing large text datasets with the use of LLMs can be performed efficiently in terms of both time and energy. For this purpose, we utilize the Llama pre-trained model. In more detail, we study the topic modeling task where the goal is to discover and identify topics in large text corpora. The basis of our approach is a hierarchical divisive clustering technique that clusters the data based on their semantic similarity, after employing a Sentence-BERT encoder, pre-trained on a variety of data across different tasks. Then, using an LLM, we identify topics for representative samples from each cluster. Additionally, we introduce a new evaluation method that leverages the capabilities of LLMs to assess the alignment between discovered topics and ground truth labels, providing a robust validation metric. Our findings indicate that it is possible to effectively reduce the computational cost of the topic modeling process compared to the direct application of LLMs and BERTopic, while simultaneously enhancing inference time and overall efficiency, thereby surpassing the current state-of-the-art capabilities of BERTopic.

Single-cell RNA sequencing (scRNA-seq) represents a paradigm shift in understanding the complexities of cellular functions and states at an individual level. Despite its transformative potential in revealing cellular heterogeneity and mechanisms, scRNA-seq data poses significant analytical challenges due to its high dimensionality, noise, and sparsity. This paper introduces Hyperdimensional Computing (HDC) as an alternative classification approach, suited for addressing these challenges. HDC, characterized by its robustness to noise and efficiency in high-dimensional spaces, accommodates the sparsity and variability inherent in scRNA-seq for the classification and analysis of scRNA-seq data offerring a promising pathway. This study aims to explore the application of HDC in scRNA-seq data classification, benchmarking its performance against traditional methods, and discussing its potential to enhance single-cell transcriptomic analysis. This work not only contributes to advancing the computational methodologies available for scRNA-seq analysis but also establishes a foundation for future research into scalable and robust data-driven approaches in genomics.

In the quest to enhance predictive models for schizophrenia spectrum disorders, Long Short-Term Memory networks (LSTM) have been pivotal due to their adept han-dling of temporal data sequences. This ability to process time-dependent data, enhances the monitoring of longitudinal patterns that are indicative of relapse. This study leverages LSTMs to analyze Heart Rate Variability (HRV), a key marker in neuropsychiatric evaluations, focusing on individual patient data. Emphasizing individualization, each LSTM model is carefully tai-lored to reflect the unique behavioral and physiological patterns of the patient. By customizing the analysis to accommodate the distinct fluctuations in HRV, LSTMs have shown a remarkable capacity to decode the complex, patient-specific patterns, as evidenced by a notable predictive average score of$0.972\pm 0.068$. This figure not only validates the LSTM model's effectiveness in grappling with the diverse aspects of psychosis but also highlights its exceptional ability to adapt to the unique temporal dynamics inherent in each patient's mental health progression.

The rapid growth of biomedical literature necessitates efficient Information Extraction systems able to identify relevant knowledge for various biological applications, such as understanding gene regulation by microRNA (miRNA). In this study, we employed a Large Language Model, specifically GPT-3.5 (version 0301), in conjunction with BERN2 for miRNA-gene interaction extraction from paper titles and abstracts. We optimized our approach using an initial dataset of about a thousand molecular biology papers and subsequently evaluated its performance on a manually curated dataset of 400 papers, achieving an accuracy of 82-85%. Driven by the promising results and the practical utility of our method, we applied the system to a large dataset of 39,000 papers. The extracted miRNA-gene interactions, combined with a Natural Language Processing approach, were included in the TarBase v9 database. Our findings demonstrate the potential of Large Language Models in biomedical Information Extraction tasks and highlight the limitations of the current gene and miRNA recognition systems, which hinder further improvements in accuracy.

Ensemble learning has been proven effective in enhancing classification accuracy by aggregating predictions from multiple base classifiers. This paper introduces a novel approach to augmenting weak projection-based classifiers using a Neural Network within a stacking ensemble framework. The proposed method capitalizes on the diverse strengths of both linear and complex models, harnessing the interpretability of projection-based classifiers, while leveraging the pattern recognition capabilities of Neural Networks. We present a comprehensive algorithm involving dataset selection, preprocessing, base model training, meta-feature generation, and Neural Network architecture design and training. Extensive experiments demonstrate the efficiency of our approach on a variety of high-dimensional biomedical datasets. Our results showcase significant accuracy improvements over standalone projection-based classifiers and conventional ensemble methods. We analyze the interpretability of the hybrid ensemble, shedding light on the insights drawn from its Neural Network component. This work not only advances the field of ensemble learning, but also underscores the potential of combining disparate classifier paradigms to achieve superior predictive performance. The code for this study is available1.1.https://github.com/panagiotisanagnostou/NNv-MRPV

The Covid-19 pandemic made a significant impact on society, including the widespread implementation of lockdowns to prevent the spread of the virus. This measure led to a decrease in face-to-face social interactions and, as an equivalent, an increase in the use of social media platforms, such as Twitter. As part of Industry 4.0, sentiment analysis can be exploited to study public attitudes toward future pandemics and sociopolitical situations in general. This work presents an analysis framework by applying a combination of natural language processing techniques and machine learning algorithms to classify the sentiment of each tweet as positive, or negative. Through extensive experimentation, we expose the ideal model for this task and, subsequently, utilize sentiment predictions to perform time series analysis over the course of the pandemic. In addition, a change point detection algorithm was applied in order to identify the turning points in public attitudes toward the pandemic, which were validated by cross-referencing the news report at that particular period of time. Finally, we study the relationship between sentiment trends on social media and, news coverage of the pandemic, providing insights into the public's perception of the pandemic and its influence on the news.

The task of understanding human behavior through intelligent systems is crucial in various domains from medical health and well-being to financial and social platforms. In this work, we propose a complete framework that takes advantage of collected sensor accelerometer data to generate a human activity behavioral model that can be supportive in predicting future development of human movement disabilities such as Osteoarthritis or even in the individual's rehabilitation after a surgery for Osteoarthritis. More precisely, we focus on estimating uncommon behaviors within daily activities as an indication for further examination. Challenge-point of the proposed methodology is the agnostic knowledge of different behaviours of individual's movement. Based on accelerometer sensor data collected from mobile devices, the proposed framework utilizes state-of-the-art Machine Learning models for Human Activity Recognition and introduces new Deep Hybrid Models for outlier detection suggesting a solid basis for further developments and wider applicability.

We are in the big data era which has affected several domains including biomedicine and healthcare. This revolution driven by the explosion of biomedical data offers the potential for better understanding of biology and human diseases. An illustrative example is the emerging single-cell sequencing technologies, which isolate and measure each cell individually, taking a step beyond the traditional techniques where consider their measurements from a bulk of cell. Although big single-cell RNA sequencing (scRNA-seq) data promises valuable insights into the cellular level, their volume poses several challenges related to the ultra-high dimensionality. Furthermore, to further elucidate the potential of these data, more insight into gene regulatory networks (GRN) is required. Network-based approaches can tackle part of the inherent complexity of human diseases, however, the challenges related to the ultra-high dimensionality are increased. Towards this direction, we propose the NIRP, an algorithm that copes with the high dimensionality of scRNA-data using a workflow based on fast multiple random projections and a radius-based nearest neighbors search. NIRP infers a gene regulatory network (GRN) from big scRNA-seq data by transforming the original data space to a lower dimensions space and capturing the similarities among gene expressions. The network is further analyzed using a random walk approach in order to achieve dense subgraphs, active to the case under study. The performance of NIRP is evaluated in a real single-cell experimental study among three well-established GRN tools. Our results make NIRP a reliable tool, able to handle big single-cell data with ultra-high dimensionality and complexity. he main advantage of this method is that it is not affected by the volume, as much as it increases, since it transforms the data space to a specific low dimensional space.

Convolutional Neural Networks (CNNs) have been established as substantial supervised methods for classification problems in many research fields. However, a large number of parameters have to be tuned to achieve high performance and good classification results. One of the most crucial parameter for the performance of a CNN is the learning rate (step) of the training algorithm. Although the heuristic search to tune the learning rate is a common practice, it is extremely time-consuming, considering the fact that CNNs require a significant amount of time for each training, due to their complex architectures and high number of weights. Approaches that integrate the adaptation of the initial learning rate in the optimization algorithm, manage to converge to high quality solutions and have been embraced by the research community. In this work, we propose an improvement of the recently proposed Adaptive Learning Rate algorithm (AdLR). The proposed learning rate adaptation algorithm (e-AdLR) exhibits excellent convergence properties and classification accuracy, while at the same time is fast and robust.

Biomedicine is undergoing a revolution driven by the explosion of biomedical data, which are generated by emerged medical imaging, sensor technologies and high-throughput technologies. An indicative example is the single cell sequencing technology which concerns the genome sequencing examination of hundreds of separate cells in a single tumor. Consequently, open challenges arising from this emerged technology and generally from the evolution of biomedical technologies under the big data perspective. Also, given the fact that approaches based on high-performance computing require high computing resources and advanced developers, solutions that reduce the problem complexity remain very attractive. Following this direction, in this paper a classification scheme based on Multiple Random Projections and Voting is presented. Random Projections offer a platform not only for a low computational time analysis by significantly reducing the data dimensionality, but also for an accurate analysis which may well exceed classical classification approaches. The proposed method was applied on real biomedical high dimensional data and compared against well-known classification schemes as to Random Projection-based cutting-edge methods. Specifically, we applied it on expression profiles for single-cell RNA-seq data from non-diabetic and type 2 diabetic human samples. Experimental results showed that based on simplistic tools we can create a computationally fast, simple, yet effective approach for biomedical Big Data analysis and knowledge discovery.

Recent sequencing technology breakthroughs have resulted in a dramatic increase in the amount of available sequencing data, enabling major scientific advances in biology and medicine. Nowadays, sequencing transcriptome data of single cells (scRNA-seq) are growing rapidly, posing new challenges in their analysis, mostly due to their high dimensionality. In this paper, we study the problem of visualizing such high-dimensional scRNA-seq data. A new visualization scheme is presented based on a customized distance matrix retrieved by applying independently Nearest Neighbors search through multiple Random Projections. The proposed method is compared against well-known dimensionality reduction and visualization techniques showing its capabilities and performance.

In the past few years, there has been a huge growth in Twitter sentiment analysis having already provided a fair amount of research on sentiment detection of public opinion among Twitter users. Given the fact that Twitter messages are generated constantly with dizzying rates, a huge volume of streaming data is created, thus there is an imperative need for accurate methods for knowledge discovery and mining of this information. Although there exists a plethora of twitter sentiment analysis methods in the recent literature, the researchers have shifted to real-time sentiment identification on twitter streaming data, as expected. A major challenge is to deal with the Big Data challenges arising in Twitter streaming applications concerning both Volume and Velocity. Under this perspective, in this paper, a methodological approach based on open source tools is provided for real-time detection of changes in sentiment that is ultra efficient with respect to both memory consumption and computational cost. This is achieved by iteratively collecting tweets in real time and discarding them immediately after their process. For this purpose, we employ the Lexicon approach for sentiment characterizations, while change detection is achieved through appropriate control charts that do not require historical information. We believe that the proposed methodology provides the trigger for a potential large-scale monitoring of threads in an attempt to discover fake news spread or propaganda efforts in their early stages. Our experimental real-time analysis based on a recent hashtag provides evidence that the proposed approach can detect meaningful sentiment changes across a hashtags lifetime.

This paper proposes a novel methodology for automatic detection and localization of gastrointestinal (GI) anomalies in endoscopic video frame sequences. Training is performed with weakly annotated images, using only image-level, semantic labels instead of detailed, and pixel-level annotations. This makes it a cost-effective approach for the analysis of large videoendoscopy repositories. Other advantages of the proposed methodology include its capability to suggest possible locations of GI anomalies within the video frames, and its generality, in the sense that abnormal frame detection is based on automatically derived image features. It is implemented in three phases: 1) it classifies the video frames into abnormal or normal using a weakly supervised convolutional neural network (WCNN) architecture; 2) detects salient points from deeper WCNN layers, using a deep saliency detection algorithm; and 3) localizes GI anomalies using an iterative cluster unification (ICU) algorithm. ICU is based on a pointwise cross-feature-map (PCFM) descriptor extracted locally from the detected salient points using information derived from the WCNN. Results, from extensive experimentation using publicly available collections of gastrointestinal endoscopy video frames, are presented. The data sets used include a variety of GI anomalies. Both anomaly detection and localization performance achieved, in terms of the area under receiver operating characteristic (AUC), were >80%. The highest AUC for anomaly detection was obtained on conventional gastroscopy images, reaching 96%, and the highest AUC for anomaly localization was obtained on wireless capsule endoscopy images, reaching 88%.

The recent technological advancements in cloud computing and the access in increasing computational power has led in undertaking the data processing derived by mobile devices. In particular, when these data are high dimensional this is indispensable, since the mobile device has to balance its processing functionalities to additional services. However, developing efficient algorithms could allow various types of analysis to be performed locally, avoiding the necessity of a constantly connected device. In this work, we present a methodology that combines lightweight dimensionality reduction and change detection techniques. The experimental results justify its impressive performance and subsequently its usefulness in several tasks.