Measuring code understandability is both highly relevant and exceptionally challenging. This paper proposes a dynamic code understandability assessment method, which estimates a personalized code understandability score from the perspective of the specific programmer handling the code. The method consists of dynamically dividing the code unit under development or review in code regions (invisible to the programmer) and using the number of revisits (NRevisit) to each region as the primary feature for estimating the code understandability score. This approach removes the uncertainty related to the concept of a "typical programmer" assumed by static software code complexity metrics and can be easily implemented using a simple, low-cost, and non-intrusive desktop eye tracker or even a standard computer camera. This metric was evaluated using cognitive load measured through electroencephalography (EEG) in a controlled experiment with 35 programmers. Results show a very high correlation ranging from rs = 0.9067 to rs = 0.9860 (with p nearly 0) between the scores obtained with different alternatives of NRevisit and the ground truth represented by the EEG measurements of programmers’ cognitive load, demonstrating the effectiveness of our approach in reflecting the cognitive effort required for code comprehension. The paper also discusses possible practical applications of NRevisit, including its use in the context of AI-generated code, which is already widely used today.

Code comprehension assessment is crucial in modern software engineering contexts, such as the emerging LLM-supported programming paradigm, where evaluating and adjusting LLM-generated code to ensure suitability, correctness, and readability is mandatory. Recent literature offers various code comprehension solutions, ranging from subjective surveys to neurophysiological-based approaches that are more personalized and operational. However, existing proposals often estimate the cognitive load experienced by programmers during code handling, using this measure as a surrogate for code comprehension. This approach has limitations: it is indirect, as other factors influence cognitive load, and a high cognitive load does not necessarily indicate a lack of code understanding. In this paper, we propose a neurophysiological and AI-based solution using a multimodal set of biosensors, including EEG and eyetracking, along with other contextual features to measure the level of code comprehension. Instead of using cognitive load as a surrogate for code comprehension, this work tackles the challenge of assessing code comprehension by employing performance-annotated ground truth. The solution is customizable, allowing adaptation to different industrial requirements, such as stringent safety and reliability needs in mission-critical software or less critical contexts. We analyze various application scenarios to minimize the intrusiveness of the solution while maintaining acceptable performance. Evaluated in a controlled experiment with 50 programmers and 7 code comprehensions tasks, the porposed solution achieved an accuracy of 69% in the prediction of correct code comprehension. This binary modelling achieve an AUC of 75%, demonstrating its viability for measuring code comprehension in modern software development. We believe that such comprehension assessment methods are essential in current “vibe coding” workflows, where AI tools assist programmers interactively, and code understanding levels must be monitored in real-time to ensure effective human-AI collaboration.

Wheezes are adventitious respiratory sounds commonly present in patients with respiratory conditions. The presence of wheezes and their time location are relevant for clinical reasons, such as understanding the degree of bronchial obstruction. Conventional auscultation is usually employed to analyze wheezes, but remote monitoring has become a pressing need during recent years. Automatic respiratory sound analysis is required to reliably perform remote auscultation. In this work we propose a method for wheeze segmentation. Our method starts by decomposing a given audio excerpt into intrinsic mode frequencies using empirical mode decomposition. Then, we apply harmonic-percussive source separation to the resulting audio tracks and get harmonic-enhanced spectrograms, which are processed to obtain harmonic masks. Subsequently, a series of empirically derived rules are applied to find wheeze candidates. Finally, the candidates stemming from the different audio tracks are merged and median filtered. In the evaluation stage, we compare our method to three baselines on the ICBHI 2017 Respiratory Sound Database, a challenging dataset containing various noise sources and background sounds. Using the full dataset, our method outperforms the baselines, achieving an F1 of 41.9%. Our method's performance is also better than the baselines across several stratified results focusing on five variables: recording equipment, age, sex, body-mass index, and diagnosis. We conclude that wheeze segmentation has not been solved for real life scenario applications. Adaptation of existing systems to demographic characteristics might be a promising step in the direction of algorithm personalization, which would make automatic wheeze segmentation clinically viable.

Code review is an essential practice in software engineering to spot code defects in the early stages of software development. Modern code reviews (e.g., acceptance or rejection of pull requests with Git) have become less formal than classic Fagan's inspections, lightweight, and more reliant on individuals (i.e., reviewers). However, reviewers may encounter mentally demanding challenges during the code review, such as code comprehension difficulties or distractions that might affect the code review quality. This work proposes a novel approach that evaluates the quality of code reviews in terms of bug-finding effectiveness and provides the reviewers with a clear message of whether the review should be repeated, indicating the code regions that may not have been well-reviewed. The proposed approach utilizes biometric information collected from the reviewer during the review process using non-intrusive biofeedback devices (e.g., smartwatches). Biometric measures such as Heart Rate Variability (HRV) and task-evoked pupillary response are captured as a surrogate of the cognitive state of the reviewer (e.g., mental workload) and inexpensive desktop eye-trackers compatible with the software development settings. This work uses Artificial Intelligence techniques to predict the cognitive load from the extracted biomarkers and classify each code region according to a set of features. The final evaluation considers various factors such as code complexity, time of the code review, the experience level of the reviewer, and other factors. Our experimental results show the approach could predict the review quality with 87.77%±4.65 accuracy and a Spearman correlation coefficient of 0.85 (p-value < 0.001) between the predicted and the actual review performance. This evaluation validates the cognitive load measurement using electroencephalography (EEG) signals as ground truth for the HRV and pupil signals.

Code reviews and software inspections are essential for building reliable software. However, current code reviews practice in the software industry (e.g., acceptance or rejection of pull requests with Git) deviates considerably from classic (and expensive) Fagan's inspections. Modern code reviews are lightweight and asynchronous and do not rely on a group of inspectors and inspection meetings any longer. The modern style of code reviews is much more flexible and cost-effective. Still, these advantages come with the price of reducing the quality of code reviews, as a single reviewer generally makes them with all the inherent and the very human limitations of one single look. The reviewer could be distracted, overloaded, under stress, or not even fully understand the code under review. This paper proposes a new tool (iReview) that evaluates the code review quality using biometric measures gathered from code reviewers (often called Biofeedback). Biometric measures such as Heart Rate Variability (HRV) and eye movement dynamics are used to assess the reviewer's comprehension of the code under review. iReview evaluates the quality of each review globally and indicates the code regions that have not been well-reviewed, explaining why those code regions should be reviewed again. The tool uses Artificial Intelligence techniques to classify the code regions into good and bad reviews based on various biometric and non-biometric features. The first results show that iReview can predict the review quality of medium or complex programs with an accuracy ranging from 75% to 87% in detecting bad reviews (i.e., code regions classified as bad reviewed still have undetected bugs). This tool is expected to improve software reliability by ensuring that good reviews have been carried out despite the current lightweight reviewing processes.

Patients with respiratory conditions typically exhibit adventitious respiratory sounds, such as wheezes. Wheeze events have variable duration. In this work we studied the influence of event duration on wheeze classification, namely how the creation of the non-wheeze class affected the classifiers' performance. First, we evaluated several classifiers on an open access respiratory sound database, with the best one reaching sensitivity and specificity values of 98% and 95%, respectively. Then, by changing one parameter in the design of the non-wheeze class, i.e., event duration, the best classifier only reached sensitivity and specificity values of 55% and 76%, respectively. These results demonstrate the importance of experimental design on the assessment of wheeze classification algorithms' performance.

Our research explores a recent paradigm called Biofeedback Augmented Software Engineering (BASE) that introduces a strong new element in the software development process: the programmers' biofeedback. In this Practical Experience Report we present the results of an experiment to evaluate the possibility of using pupillography to gather biofeedback from the programmers. The idea is to use pupillography to get meta information about the programmers' cognitive and emotional states (stress, attention, mental effort level, cognitive overload,...) during code development to identify conditions that may precipitate programmers making bugs or bugs escaping human attention, and tag the corresponding code locations in the software under development to provide online warnings to the programmer or identify code snippets that will need more intensive testing. The experiments evaluate the use of pupillography as cognitive load predictor, compare the results with the mental effort perceived by programmers using NASATLX, and discuss different possibilities for the use of pupillography as biofeedback sensor in real software development scenarios.

Recent studies have shown that programmers' cognitive load during typical code development activities can be assessed using wearable and low intrusive devices that capture peripheral physiological responses driven by the autonomic nervous system. In particular, measures such as heart rate variability (HRV) and pupillography can be acquired by nonintrusive devices and provide accurate indication of programmers' cognitive load and attention level in code related tasks, which are known elements of human error that potentially lead to software faults. This paper presents an experimental study designed to evaluate the possibility of using HRV and pupillography together with eye tracking to identify and annotate specific code lines (or even finer grain lexical tokens) of the program under development (or under inspection) with information on the cognitive load of the programmer while dealing with such lines of code. The experimental data is discussed in the paper to assess different alternatives for using code annotations representing programmers' cognitive load while producing or reading code. In particular, we propose the use of biofeedback code highlighting techniques to provide online programmer's warnings for potentially problematic code lines that may need a second look at (to remove possible bugs), and biofeedback-driven software testing to optimize testing effort, focusing the tests on code areas with higher bug probability.

Neurally mediated syncope (NMS) patients suffer from sudden loss of consciousness, which is associated with a high rate of falls and hospitalization. NMS negatively impacts a subject's quality of life and is a growing cost issue in our aging society, as its incidence increases with age. In this paper, we present a solution for prediction of NMS, which is based on the analysis of the electrocardiogram (ECG) and photoplethysmogram (PPG) alone. Several parameters extracted from ECG and PPG, associated with reflectory mechanisms underlying NMS in previous publications, were combined in a single algorithm to detect impending syncope. The proposed algorithm was evaluated in a population of 43 subjects. The feature selection, distance metric selection, and optimal threshold were performed in a subset of 30 patients, while the remaining data from 13 patients were used to test the final solution. Additionally, a leave-one-out cross-validation scheme was also used to evaluate the performance of the proposed algorithm yielding the following results: sensitivity (SE)--95.2%; specificity (SP)--95.4%; positive predictive value (PPV)--90.9%; false-positive rate per hour (FPRh)-0.14 h(-1), and prediction time (aPTime)--116.4 s.

This paper aims to assess the predictive value of physiological data daily collected in a telemonitoring study in the early detection of heart failure (HF) decompensation events. The main hypothesis is that physiological time series with similar progression (trends) may have prognostic value in future clinical states (decompensation or normal condition). The strategy is composed of two main steps: a trend similarity analysis and a predictive procedure. The similarity scheme combines the Haar wavelet decomposition, in which signals are represented as linear combinations of a set of orthogonal bases, with the Karhunen-Loève transform, that allows the selection of the reduced set of bases that capture the fundamental behavior of the time series. The prediction process assumes that future evolution of current condition can be inferred from the progression of past physiological time series. Therefore, founded on the trend similarity measure, a set of time series presenting a progression similar to the current condition is identified in the historical dataset, which is then employed, through a nearest neighbor approach, in the current prediction. The strategy is evaluated using physiological data resulting from the myHeart telemonitoring study, namely blood pressure, respiration rate, heart rate, and body weight collected from 41 patients (15 decompensation events and 26 normal conditions). The obtained results suggest, in general, that the physiological data have predictive value, and in particular, that the proposed scheme is particularly appropriate to address the early detection of HF decompensation.

We introduce a context-aware keypoint extractor, coined as CAKE, aimed at capturing the most informative image content. We find this algorithm particularly useful in tasks such as image retrieval, scene classification, and object (class) recognition, in which local features are mainly used to provide a robust and efficient image representation. We are motivated by the fact that the majority of local feature extractors are designed to respond to a reduced number of structures. Furthermore, we observe that the existent complementarity among feature sets is often neglected. Our context-aware algorithm is designed to respond to complementary features as long as they are informative. In the particular case of images with different types of structures, one can expect a high complementarity among the features retrieved by a context-aware extractor. By contrast, images with repetitive patterns will inhibit our method from retrieving a clear summarised description of the image content. Nonetheless, the extracted set of features can be complemented with a counterpart that retrieves the repetitive elements in the image. These two cases are depicted in Figure 1. The upper image shows a context-aware keypoint extraction on a well-structured scene, which retrieves the 100 most informative keypoints. This small number of features is sufficient to provide a good coverage of the content, which includes different types of structures. The lower image illustrates the advantages of combining context-aware keypoints with strictly local ones (SFOP keypoints [2]) to obtain a better coverage of images with repetitive patterns. An information theoretic framework is used to formulate our contextaware keypoint extraction. A keypoint will correspond to a certain image location within a structure with a low probability of occurrence (high information content). For each image location x, we consider w(x) ∈RD, any viable local representation (e.g, the Hessian matrix or the structure tensor matrix) as a “codeword” that represents the neighbourhood of x. To define the saliency measure, we regard the image codewords as samples of a multivariate probability density function. We compute the probability of a codeword w(y) using a Kernel Density Estimator [4] in which the kernel is a multidimensional Gaussian function with zero mean and standard deviation σk:

Heart sound is one of the significant bio-signals to diagnose certain cardiac anomalies. Aiming to provide an automatic heart sounds analysis to medical professionals, we present a method for heart sound segmentation based on the signal's simplicity and strength. The method has two phases: the first phase identifies the timings of SI and S2 sound limits using simplicity and strength in the wavelet domain; the second phase discriminates among the SI and S2 using high frequency information. The method incorporates an adaptive approach for thresholding the simplicity and strength profiles to detect the timings of the sound in the presence of murmur. The algorithm was tested using a database that contains 7 types of murmurs of various grades. An average accuracy of 4.22 ms was achieved in SI and S2 boundaries identification. Regarding SI and S2 classification overall sensitivity and specificity of 91.78% and 90.41% were obtained, respectively.

In this work, we propose a two-stage classifier based on the analysis of the heart sound's complexity for murmur identification and classification. The first stage of the classifier verifies if the heart sound (HS) exhibits murmurs. To this end, the chaotic nature of the signal is assessed using the Lyapunov exponents (LEs). The second stage of the method is devoted to the classification of the type of murmur. In opposition to current state of the art methods for murmur classification, a reduced set of features is proposed. This set includes both well-known as well as new features designed to capture the morphological and the chaotic nature of murmurs. The classification scheme is evaluated with three classification methods: Learning Vector Quantization, Gaussian Mixture Models and Support Vector Machines. The achieved results are comparable to reported results in literature, while relying on a significant smaller set of features.

A comparison and a performance evaluation of 3D interest point detectors based on the structure tensor matrix under a medical image registration context is presented. The study regards the distinctiveness and the repeatability rate exhibited by the detectors on medical images as the fundamental criteria for selecting a detector to be the support of a registration task. We empirically assess those requirements under the specificities of medical image analysis and registration.

Atrial fibrillation (AF) is an arrhythmia that can lead to several patient risks. This kind of arrhythmia affects mostly elderly people, in particular those who suffer from heart failure (one of the main causes of hospitalization). Thus, detection of AF becomes decisive in the prevention of cardiac threats. In this paper an algorithm for AF detection based on a novel algorithm architecture and feature extraction methods is proposed. The aforementioned architecture is based on the analysis of the three main physiological characteristics of AF: i) P wave absence ii) heart rate irregularity and iii) atrial activity (AA). Discriminative features are extracted using model-based statistic and frequency based approaches. Sensitivity and specificity results (respectively, 93.80% and 96.09% using the MIT-BIH AF database) show that the proposed algorithm is able to outperform state-of-the-art methods.

This paper presents a new algorithm for segmentation and classification of S1 and S2 heart sounds without ECG reference. The proposed approach is composed of three main stages. In the first stage the fundamental heart sound lobes are identified using a fast wavelet transform and the Shannon energy. Next, these lobes are validated and classified into S1 and S2 classes based on Mel-frequency coefficients and on a non-supervised neural network. Finally, regular heart cycles are identified in a post-processing stage by a heart rhythm criterion. This approach was tested using sound samples collected from prosthetic valve implanted patients. Results are comparable with ECG based approaches

Spectral data estimation from image data is an ill-posed problem since (i) due to the integral nature of imaging sensors, the same output can be obtained from an infinity of input signals and (ii) color signals are spectrally smooth in nature and, therefore, limit the number of linear independent data than can be collected. To enable the solution of these problems the solution's search space has to be constrained. The question that arises is how to select/parameterize these constraints? In this paper several model selection criteria are extended for spectral data estimation and evaluated in the context of spectral sensitivity function estimation of CCD sensors.

A robust image watermarking scheme relying on an affine invariant embedding domain is presented. The invariant space is obtained by triangulating the image using affine invariant interest points as vertices and performing an invariant triangle representation with respect to affine transformations based on the barycentric coordinates system. The watermark is encoded via quantization index modulation with an adaptive quantization step

Recovering spectral sensitivities of imaging devices with indirect methods, as well as spectral stimuli estimation from device responses are ill-posed problems. All known methods have to rely on a priori information to constrain the solution space, which, in most situations, is difficult or even impossible to obtain. In this paper we introduce a simple and fully data-driven approach for indirect spectral sensitivity estimation, which does not rely on explicit a priori information. The method is built upon an extension of our previous work on generalized cross-validation for constraint Tikhonov problems and utilizes a linear combination of band-limited basis functions.

Light sensor spectral calibration is an ill-defined problem. For the identification problem one needs a priori knowledge of the characteristics of the sensor which is difficult to get in most situations. A new methodology is presented in this paper that does not rely on any a priori knowledge of the sensor's characteristics. The method uses an extended generalized cross-validation function to measure predictability of the identified sensor's spectral behavior. The prediction error is minimized with a hybrid genetic algorithm. Further an extended image formation model is introduced to model changes in additive and multiplicative errors. The calibration problem is formulated to be independent of these changes by previously identifying and removing them from the images.

Spectral data estimation from image data is an ill-posed problem since (i) due to the integral nature of solid-state light sensors, the same output can be obtained from an infinity of input signals and (ii) color signals are spectrally smooth in nature and therefore limit the number of linear independent equations that can be formulated for the identification problem. To enable the solution of these problems most methods rely on exact a priori knowledge, such as smoothness and modality, to formulate hard constraints. A new method based on an extended generalized cross-validation measure is introduced for this type of problems. The solution is obtained with a genetic algorithm that maximizes its prediction ability. The method does not require exact a priori knowledge on the solution, since it is able to extract this information from the input data.