The stakes are far from academic. During 2011 and 2012, droughts across the central United States caused more than 30 million dollars in agricultural losses, and rapid-onset flash droughts, which intensify in a matter of weeks, have become an increasing concern for farmers, water managers, and emergency planners. Effective monitoring and detection tools are essential for managing these risks, yet most existing approaches to drought mapping share a fundamental weakness: they treat space and time as separate dimensions. A typical analysis first identifies spatial clusters of dry conditions at a given moment, then stitches those snapshots together across time. Because space and time are genuinely inseparable in a physical drought, that separation introduces error at every step, blurring the boundaries of events and sometimes merging distinct droughts into one another.

The new method, developed by T. Elana Christian and colleagues, tackles this problem head-on by treating space and time jointly at every stage of the analysis. The algorithm operates in two levels. The first level is a modified version of the classic k-means clustering algorithm, a workhorse of data science that partitions data points into a user-defined number of groups. In standard k-means, each data point is assigned to the cluster whose centroid, or center, is closest by some distance measure. When applied naively to a drought index, this produces clusters stratified purely by intensity, with locations flickering rapidly between drought states from one day to the next, a pattern the authors describe as nonphysical. Real landscapes do not transition between drought conditions instantly.

The innovation lies in how the modified algorithm defines closeness. Rather than measuring the distance between isolated space-time data points, the researchers first construct a spatiotemporal neighborhood around every point, consisting of adjacent locations and days whose values are correlated with the central point. The assignment step then averages distances across this neighborhood, so that a point’s cluster membership depends not just on its own intensity but on the intensity of its surrounding space-time context. The result is clustering based on both intensity and local spatiotemporal consistency. When the team compared the two approaches on data from 2003, the standard k-means produced erratic, rapidly shifting drought regimes, while the space-time version generated temporally persistent regimes with realistic durations and transitions that mirrored the physical evolution of drought.

Practical considerations shaped the implementation as well. The dataset is enormous: daily values of a drought index across the continental United States at a resolution of roughly 14 kilometers, spanning 42 years from 1980 to 2021. To handle data of this scale, the researchers employed a minibatched variant of k-means, which optimizes the clustering objective over random batches of data rather than the full dataset, trading a small amount of precision for dramatically faster convergence. The drought index itself, called the standardized vapor pressure deficit drought index, or SVDI, was introduced in 2022 as a simplified, temperature-based measure of atmospheric evaporative demand, calculated from daily maximum air temperature and minimum relative humidity. Higher SVDI values indicate drier conditions, and the index has previously proven capable of identifying known flash droughts and large-scale drought variability.

Between the two levels of the algorithm sits an optional, expert-informed step. After the space-time k-means clusters are formed, their mean values are computed and classified into drought severity categories, following conventions similar to those used for the standardized precipitation evapotranspiration index. Clusters whose means exceed 0.5, corresponding to mild, moderate, severe, or extreme drought, are selected and passed forward, while clusters in the same severity category are collapsed together. This intermediate filtering reduces the computational load on the second level and sharpens its performance. The second level then applies DBSCAN, a density-based clustering algorithm from 1996 that groups points lying in high-density regions and labels isolated points as noise. Crucially, DBSCAN does not require the number of clusters to be specified in advance and can carve out clusters of arbitrary shape, making it well suited to separating drought events that share similar statistical properties but occupy different regions of space and time.

The researchers tuned the full pipeline, which has six tunable parameters in total, against the well-documented summer 2003 flash drought in the central United States, an event marked by rapid intensification beginning in late June and a rapid demise before September. Spatiotemporal neighborhoods were sized according to the correlation structure of the data: temporal autocorrelation decayed within about a week, so a 14-day window was chosen, while spatial correlations became negligible beyond roughly 127 kilometers. The number of k-means clusters was set to five, and the DBSCAN neighborhood radius was selected by inspecting how the extracted clusters changed across a range of values. Too small a radius produced unrealistic, intensity-stratified clusters; too large a radius merged distinct events into one. The chosen value cleanly separated three distinct drought structures in 2003, including a large event spanning the southwestern and central United States, and the algorithm’s output aligned closely with the known spatial extent and timing of the historical drought.

A key test of robustness came when the team applied parameters tuned on 2003 to an entirely different year, 2000, which featured a flash drought originating in Florida, Mississippi, and Georgia alongside a longer-term drought across Texas, the Plains, and the Southwest. Even without retuning, the algorithm captured the southeastern flash drought with reasonable accuracy, briefly conflating it with the contemporaneous Texas and Plains drought for about two weeks before recovering the correct location. When the DBSCAN parameters were tuned specifically to 2000, the misidentification disappeared entirely, and the drought was classified as extreme rather than severe. The authors attribute the residual errors in the transfer test to the spatial smoothness of the drought index itself, which can link regions experiencing different events through continuously high values.

Running the algorithm year by year across the full 1980 to 2021 record revealed long-term shifts in American drought behavior. Mild and moderate droughts dominated the first half of the period, while severe and mild droughts predominated in the latter half. The frequency of mild droughts has increased markedly: between 1981 and 2000 there were years with no mild drought at all, but from 2000 to 2021 a mild drought occurred almost every year. Drought durations are changing too. After 2000, mild droughts became more likely to persist for an entire year, severe droughts are lasting longer and occurring more frequently than moderate ones, and the longest extreme drought in the dataset occurred in 2011. These trends correspond to known patterns in the drought literature, lending independent credibility to the method’s output.

Beyond drought, the framework is deliberately generalizable. Because the space-time width of the neighborhoods and the DBSCAN parameters can be adjusted, the method can be tuned to events at very different scales, from short-lived flash droughts to multi-season megadroughts, and the authors suggest it could be applied to other spatiotemporally continuous environmental phenomena. They also note that either level of the pipeline could be swapped for alternative algorithms, such as HDBSCAN, BIRCH, or Voronoi-based clustering, each offering advantages in speed or the ability to handle clusters of varying density. The code has been released publicly on GitHub, and the team highlights a remaining challenge common to unsupervised approaches: automating parameter tuning toward specific environmental phenomena, so that domain expertise, while still valuable, is no longer a prerequisite for turning raw climate data into a catalog of discrete, traceable events.

Subject of Research: A bi-level spatiotemporal clustering algorithm for extracting drought events from a vapor pressure deficit drought index over the continental United States

Article Title: A bi-level spatiotemporal clustering approach and its application to drought extraction

Article References: Christian, T. E., Subrahmanya, A. N., Gamelin, B., Rao, V., Samia, N. I., & Bessac, J. (2025). A bi-level spatiotemporal clustering approach and its application to drought extraction. Advances in Statistical Climatology, Meteorology and Oceanography, 11(2), 257-272. https://doi.org/10.5194/ascmo-11-257-2025

Image Credits: AI Generated

DOI: 10.5194/ascmo-11-257-2025

Keywords: drought, spatiotemporal clustering, k-means, DBSCAN, flash drought, vapor pressure deficit, SVDI, machine learning, climate extremes, United States, NLDAS, drought monitoring

The method, called Alignment-Based Deformation Modeling, or ABDM, was developed and validated by Thanh Bao Nhut Trinh and Tsung-Chin Hou, who subjected three full-scale reinforced concrete beams, each four meters long, to three-point bending tests in the laboratory. Instead of relying on point-wise contact instruments such as linear variable differential transformers, the team scanned each beam with a terrestrial laser scanner at every loading stage, from the unloaded reference state through to failure. The scanner, a Leica ScanStation C10, captured dense three-dimensional point clouds of the entire beam surface, providing the full-field geometric data that conventional sensors cannot offer.

The technical core of the approach lies in how it processes those clouds. Rather than comparing raw point sets directly—a strategy the authors argue is fragile under large deformation, cracking, or partial data loss—the framework first reconstructs a structured parametric geometric model of the unloaded beam. Using the efficient RANSAC algorithm, it segments the point cloud into the planar faces of the prismatic beam. Because laser scans inevitably suffer from occlusion, particularly on rear faces hidden from the scanner, the workflow then infers the missing surfaces: it classifies the connections between adjacent planes as convex or concave based on their surface normals and centroids, and generates the absent faces by translating detected surfaces along their outward normals by the beam width. The result is a closed geometric envelope representing the complete as-built beam.

From that envelope, the pipeline extracts boundary points using Delaunay triangulation combined with statistical neighborhood filtering, then slices the edge data into segments along the beam axis. Within each slice, the DBSCAN clustering algorithm groups boundary points into the corners of the beam cross-section, and each cluster is turned into a rectangular prism whose dimensions come directly from the measured geometry. Merging these volumetric slices yields a parametric model of the undeformed beam—a reference against which every subsequent loading stage can be judged. The authors are careful to note that the term 3D model here means a structured parametric geometric representation carrying only geometric attributes, not a semantically enriched building information model with materials or scheduling data.

Modeling the loaded beam is where the alignment strategy earns its name. Loaded scans are registered to the unloaded reference rather than to isolated undeformed features, because under three-point bending the beam’s surface features themselves deform, crack, and rotate as the test progresses. The workflow slices both datasets around the expected cracking zone, rotates the loaded segments to align with the global axis, and applies Iterative Closest Point registration to each pair. When large cracks or local rotations cause a single rigid registration to become unstable, the algorithm splits the loaded data along failure zones and re-runs the registration independently on each sub-segment, iterating until the root mean square error falls below a threshold of 0.01. The recovered transformation matrices are then applied to the corresponding parametric segments of the reference model, mapping them into the deformed configuration and producing a complete as-loaded geometric model.

The framework also detects surface spalling—material loss on the beam face—by identifying holes and abrupt discontinuities in the extracted front surface of the loaded scan. Spalling area is computed as the difference between the reconstructed full front surface area and the area of the extracted damaged region, using an alpha-shape method that handles irregular boundaries. In the experiments, Beam A, cast from conventional 280 kgf/cm² reinforced concrete, developed a spalling zone of roughly 0.028 square meters near midspan at the final loading stage. Beam B, which incorporated steel fibers in its midspan region, showed only about 0.00184 square meters—a reduction of approximately 90 percent—because the fibers bridge cracks and resist tensile failure. Beam C, reinforced with polyvinyl-alcohol fibers, exhibited no detectable spalling at all, consistent with the tight crack-width control that polymeric fibers provide.

Validation against independent reference measurements demonstrated millimeter-level accuracy. Linear variable differential transformers installed at midspan and the quarter points recorded vertical displacements throughout each test, and the automated model-based deflections tracked them closely. Mean deflection errors of the ABDM method relative to the LVDTs stayed within about ±1 millimeter, or roughly ±2 percent, with relative errors never exceeding 3.60 percent. Bland–Altman analysis showed mean biases below 1 millimeter for all three beams—0.72 millimeters for Beam A, −0.19 for Beam B, and 0.03 for Beam C—with more than 95 percent of differences falling within narrow limits of agreement. A comparison against a least-squares curve-fitting approach showed the new method to be more stable, particularly for Beam A, where wide, localized cracking and lateral torsional movement made polynomial fitting err by nearly 4 millimeters, about 10 percent, at midspan.

Speed is the framework’s headline claim. Once the algorithm parameters—RANSAC thresholds, DBSCAN clustering radius, and ICP convergence tolerance—were fixed for the point cloud density, the entire workflow ran without manual intervention from raw scan to deformation analysis. On an ordinary desktop computer with an Intel Core i7-8700 processor and 16 gigabytes of RAM running MATLAB, generating the 3D models and computing deformations took between 40 and 48 seconds per beam, with the final deformation calculation itself requiring only about 2 seconds. The authors position the method as a rapid automated post-processing tool rather than a strict real-time monitoring system, but the contrast with conventional Scan-to-BIM workflows, which typically demand semi-automatic region selection and interactive alignment, is stark.

The study is candid about its limits. The geometric inference assumes prismatic beams with planar, roughly orthogonal faces, so non-prismatic members, diagonal elements, and complex joints would require relaxed constraints and additional segmentation. The alignment relies exclusively on surface-based registration without external control targets, and the spalling quantification was compared against manual delineation of the same point clouds rather than an independent ground truth, yielding relative errors below 6.5 percent and Intersection over Union values up to 0.88. All three specimens were laboratory beams tested under controlled conditions, and the authors acknowledge that real-world structures with complex geometries, environmental variability, and practical scanning constraints remain untested territory.

Even so, the implications reach well beyond the laboratory. Because the method needs no training data, it sidesteps the large annotated datasets and computational demands that constrain deep-learning approaches to damage detection, while offering interpretability that black-box models lack. It also complements techniques such as three-dimensional digital image correlation, which achieves sub-millimeter precision but requires controlled lighting, textured surfaces, and a limited field of view. The authors point toward seismic performance evaluation and routine monitoring of reinforced concrete components as near-term applications, with future work targeting the computational efficiency needed for large datasets from bridges and high-rise buildings. If automated pipelines like this one mature, the era of inferring a structure’s health from a few glued-on gauges may give way to full-field, model-based assessment delivered in under a minute.

Subject of Research: Automated point cloud-to-BIM conversion for deformation and spalling monitoring of reinforced concrete beams

Article Title: Automated point cloud-to-BIM conversion for fast deformation monitoring of structural elements

Article References: Trinh, T. B. N., & Hou, T.-C. (2026). Automated point cloud-to-BIM conversion for fast deformation monitoring of structural elements. Results in Engineering, 32, Article 113270. https://doi.org/10.1016/j.rineng.2026.113270

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113270

Keywords: terrestrial laser scanning, point cloud, BIM, structural health monitoring, deformation monitoring, reinforced concrete, three-point bending, spalling detection, ICP registration, RANSAC, DBSCAN, deflection measurement

The appeal of so-called frozen orbits is well established. These are orbits whose size, shape, and orientation remain nearly constant in a frame rotating with the Moon, requiring almost no propellant to maintain. The Lunar Reconnaissance Orbiter demonstrated the payoff: by carefully selecting a polar frozen orbit, the mission used only 33.9 kilograms of hydrazine over twelve years for all of its station-keeping, phasing, and momentum management. Frozen orbits could similarly support extended surface observation campaigns, communications relays, and even the tracking of orbital debris or derelict hardware from past missions, such as the Apollo-era lunar module ascent stages that may still circle the Moon.

Traditionally, frozen orbits have been found analytically. Researchers construct an averaged dynamical model that strips out short-period oscillations, then solve for equilibrium points in the averaged equations. These classical analyses predict that frozen orbits occur under specific conditions: when the mean argument of perilune, the point of closest approach, sits near plus or minus 90 degrees; when the mean eccentricity is zero; or when the argument of perilune equals 0 or 180 degrees. The trouble is that averaging discards real physics, and the resulting predictions must be verified by propagating trajectories in a high-fidelity model that includes a detailed lunar gravity field and the true ephemerides of the Earth and Sun. Scaling that verification across the whole phase space has always been the bottleneck.

Bosanac and Miceli attacked the bottleneck directly. They defined over 1.1 million initial conditions at a fixed epoch, sampling orbital elements coarsely across semi-major axes up to 2,038 kilometers, eccentricities, inclinations from 0 to 180 degrees, and the full range of nodal longitudes and arguments of perilune. Each initial state was propagated for up to 180 days in a demanding dynamical model: a 100 by 100 degree-and-order spherical harmonic lunar gravity field, drawn from the GRGM900C model derived from GRAIL data, combined with the point-mass gravity of the Earth and Sun positioned according to the DE421 ephemerides. Integrations used a high-order Runge-Kutta scheme with tight tolerances of ten to the minus eleventh, terminating early if the spacecraft struck a spherical Moon of radius 1,738 kilometers.

Of the initial conditions tested, 638,181 trajectories survived at least three lunar rotational periods, roughly 81.9 days, and entered the analysis. Rather than storing every point along every path, the researchers represented each trajectory by its sequence of perilunes, the moments of closest approach, which capture the long-period evolution of the orbit’s eccentricity and orientation. Each perilune was described by two compact coordinates, p and q, equal to the eccentricity times the cosine and sine of the argument of perilune. These coordinates encode both the shape of the orbit and the direction in which its low point points, and they trace out distinctive loops and curves as the orbit evolves in the Moon-fixed frame.

The clustering itself relied on DBSCAN, a density-based algorithm that groups points lying in sufficiently crowded neighborhoods while labeling isolated points as noise. Because the dataset was too large to process at once, it was split into 64 partitions of roughly 10,000 trajectories each. Within each partition, trajectories were grouped if their perilune sequences moved through the pq-plane along similar paths at similar times, judged by unit tangent vectors sampled every half day. Local clusters were then merged across partitions and across different initial phasings, using shape-based feature vectors and a graph structure whose connected components defined the final groupings. The outcome was striking: 638,181 trajectories collapsed into just 885 global clusters, a reduction of three orders of magnitude in the information an analyst must examine.

From selected clusters, the team manually extracted sets of long-lifetime orbits, defined as trajectories showing limited drift in the pq-plane over the full 180 days. These orbits appear at discrete combinations of initial semi-major axis and inclination, and in many cases, shifting the initial longitude of the ascending node rotates the entire groundtrack geometry without changing the underlying orbital evolution. At a semi-major axis of 1,808 kilometers and an inclination of 85 degrees, for example, three different nodal longitudes produced long-lifetime orbits whose perilune groundtracks circled the north pole, circled the south pole, or swept across both hemispheres, all while the eccentricity and argument of perilune evolved in nearly identical fashion. That flexibility matters for mission design, because the groundtrack determines which terrain a spacecraft observes.

The comparison with classical theory was largely reassuring, with intriguing exceptions. Most of the recovered long-lifetime orbits cluster around arguments of perilune near plus or minus 90 degrees or encircle zero eccentricity in the pq-plane, exactly as analytical frozen orbit theories predict. The overall shapes of the mean eccentricity curves as a function of inclination also match the families derived by Lara in 2011 and later refined by Yarndley and colleagues. But at initial inclinations of 65 and 115 degrees, the numerical search surfaced low-drift orbits with mean arguments of perilune as far as 30 to 60 degrees away from the predicted minus 90 degrees. When the authors extended two of these trajectories to nearly twenty years, some remained tightly bounded, hinting at frozen orbits that the averaged theories miss, while others drifted substantially under longer-period oscillations governed by multiple frequencies.

The method also recovered orbits that analytical approaches cannot touch by construction. Resonant trajectories, whose perilune sequences form looping paths in the pq-plane, are typically excluded from averaged derivations, yet the clustering search found them naturally at low inclinations. Several results echoed earlier numerical work: orbits at 95 degrees inclination showed oscillation periods of 27.43 to 27.54 days, close to the 27.35-day repeat groundtrack period reported by Russell and Lara in 2007, and near-equatorial long-lifetime orbits resembled those designed for the DARE radio astronomy concept by Plice and colleagues.

The authors are careful to frame the work as a proof of concept. The initial conditions were sampled coarsely, at a single epoch, and only over 180 days, so some families may be missing and some orbits may not persist indefinitely. Extracting representative orbits from clusters is still manual, and automating that step is ongoing work. Even so, the demonstration is compelling: a data-mining pipeline can compress a million brute-force simulations into a compact, interpretable summary that agrees with decades of theory while exposing solutions theory overlooked. As NASA and its partners plan sustained lunar operations under Artemis, tools like this could quietly become as important as the rockets, ensuring that the spacecraft placed around the Moon stay there without burning fuel they cannot spare.

Subject of Research: Numerical identification of long-lifetime low lunar orbits using trajectory clustering in a high-fidelity lunar gravity model

Article Title: Numerical exploration of low lunar, long-lifetime orbits

Article References: Bosanac, N., & Miceli, G. E. (2026). Numerical exploration of low lunar, long-lifetime orbits. Celestial Mechanics and Dynamical Astronomy, 138(5), Article 61. https://doi.org/10.1007/s10569-026-10334-x

Image Credits: AI Generated

DOI: 10.1007/s10569-026-10334-x

Keywords: lunar orbits, frozen orbits, trajectory clustering, celestial mechanics, lunar gravity field, mission design, DBSCAN, perilune, space situational awareness, astrodynamics, Lunar Reconnaissance Orbiter, ephemeris model

The paper’s most striking number concerns how these pipelines have been measured, not how they perform. When feature selection, scaling, outlier removal and the popular SMOTE-ENN resampling technique are all fitted on the full dataset before cross-validation — a practice the authors trace through much of the published literature — reported F1 scores are inflated by an average of 28.9 percentage points, with a range from 0.4 to 61.2 points depending on the cohort. The dominant culprit is resampling before splitting: SMOTE synthesises new minority-class points by interpolating between existing ones, so a test-fold patient can become a near-duplicate of training data, while the accompanying Edited Nearest Neighbours rule deletes borderline majority records using their labels. The evaluation, in effect, grades the model on questions it has already seen.

The inflation is largest precisely where the underlying task is hardest. On a small hypertension cohort of 175 patients, where blood pressure is essentially unpredictable from anthropometric measurements alone, the flawed protocol reports an F1 of 93.75 percent; the corrected protocol yields 35.5 percent, with the area under the ROC curve hovering near chance. On an easy, well-posed chronic kidney disease diagnosis task, the difference between protocols is a mere 0.4 points. Leakage, the authors conclude, does not add a constant bonus — it manufactures performance in proportion to how little genuine signal exists, which is why mid-to-high-nineties accuracy figures on small clinical cohorts deserve deep suspicion.

Under the corrected protocol, with every fitted step confined inside the training fold, the benchmark’s verdict on outlier removal is unambiguous. Twelve detectors — including tuned DBSCAN, Local Outlier Factor, Isolation Forest, One-Class SVM, robust covariance estimation, HDBSCAN, and the recent distribution-based methods ECOD and COPOD — were compared against a no-cleaning control over 50 folds per cell, using the Nadeau–Bengio corrected resampled t-test with Holm adjustment to account for overlapping training sets. Across all 130 method-by-cohort cells, no detector significantly beat the control. The single statistically significant result in either direction was a loss: HDBSCAN on the diabetic retinopathy cohort, dropping F1 by 9.37 points.

The authors did not merely tune their own detectors carefully and declare victory. To test the objection that poor configuration explains the null result, they granted an oracle: for each cohort, the best of 44 parameter settings of their proposed method was selected with hindsight on the very folds being reported. Even this impossible advantage yielded a mean gain of only 1.24 points — less than the 3.20 points that picking the best of 44 genuinely equivalent settings would produce from fold-to-fold noise alone. On two cohorts, the oracle’s chosen configuration removed nothing whatsoever. Across all 440 evaluated configurations, F1 correlated negatively with the fraction of data removed on seven of nine informative cohorts, as strongly as r = −0.99. Within this family of methods, removing more is monotonically worse.

If cleaning were merely useless, it would be a wasted step. The study’s forensic analysis shows it is actively hazardous. Every detector examined removes minority-class records preferentially — the proposed YadroSeg method deleted 32.8 percent of minority-class patients against 18.5 percent of majority-class records, a pattern significant on 8 of 10 cohorts, while Isolation Forest showed the most extreme ratio at 4.4 to 1. Triaging the deleted rows against physiological ranges, only about 15 percent contained impossible values that could be called genuine data errors; roughly 63 percent were rare but clinically plausible patients. The mechanism is not a defect in any particular algorithm: removal correlates strongly with local sparsity in feature space (r between 0.43 and 0.69) and barely at all with label impurity. Density-based detectors find sparse regions, and in clinical data sparse regions are populated by uncommon patients, not corrupted records.

The study is also a remarkable act of scientific self-correction. An earlier version of this work reported that its graph-based detector, YadroSeg, achieved the best F1 on three of five datasets, reaching 98.87 percent — a figure produced by the very leakage the revised paper dissects. The same configuration scores 100.00 under the flawed protocol and 58.51 under the corrected one. The authors pre-registered a series of tests to understand why their density-variation analysis, adapted from image segmentation, failed to transfer to tabular data, and report that two of their own registered explanations were refuted by the experiments designed to check them. The final finding is structural: driving the coring threshold until no core vertex survives leaves the set of removed records bit-identical on all ten cohorts, because the density sequence over a constructed clinical feature graph lacks the small set of dominant boundaries that thresholding presupposes.

What survives the transfer is the density ordering itself. When exposed as a continuous score rather than collapsed into a threshold, the peeling sequence recovered injected corruption better than the published algorithm on ten of ten cohorts, ranking second of nine scorers behind LOF. Measuring this honestly required the authors to correct their own evaluation a second time: their initial comparison scored detectors on the arbitrary masks they emitted, at flag rates differing by more than an order of magnitude. At a matched flag budget, every detector tested ranked corrupted records two to three times above chance — though none exceeded a detection F1 of 0.275, confirming that none of them is a reliable corruption detector, particularly for flipped labels, which leave a record exactly where it was in feature space.

The corrected implementation also brings a scalability lesson. The original formulation required a dense distance matrix that scaled as N to the power 2.67 and became infeasible beyond 32,000 records; the revised version, using radius queries on a spatial index, measures at N to the power 1.48 — competitive with DBSCAN and LOF, though an order of magnitude behind Isolation Forest’s effectively flat profile. The authors withdraw their earlier claim of O(N log N) scaling as unsupported by measurement, and confirm the exponents on two real clinical registries of over 100,000 and 250,000 records. A large-cohort confirmation arm on those registries, reaching nearly nine times the sample size of the largest primary cohort, reproduced the central null: cleaning never helped, and where differences reached significance they ran against it.

The practical recommendations are pointed. Every fitted preprocessing step should live inside the cross-validation loop, ideally enforced structurally through pipeline tools rather than manual sequencing. Any study adopting outlier removal should include a no-cleaning control and report the class composition of what it deletes. Calibration should be reported alongside discrimination, since resampling is known to damage probability calibration — a property the study observed across all configurations. And for reviewers, the single most informative check on a paper using the cleaning-plus-resampling-plus-ensemble template is whether the resampler is fitted inside or outside the validation loop. On the balance of evidence, the authors recommend against routine density-based outlier removal in clinical tabular prediction: a measurable risk of discarding rare patients, with no measurable benefit in return.

Subject of Research: Evaluation leakage and outlier removal in machine-learning prediction of chronic diseases from clinical tabular data

Article Title: Graph-based density-variation outlier removal for chronic disease prediction: A leakage-free benchmark and component-wise ablation

Article References: Davronov, R. R., & Adilova, F. T. (2026). Graph-based density-variation outlier removal for chronic disease prediction: A leakage-free benchmark and component-wise ablation. Results in Engineering, 32, Article 112842. https://doi.org/10.1016/j.rineng.2026.112842

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.112842

Keywords: machine learning, outlier detection, data leakage, chronic disease prediction, SMOTE, class imbalance, DBSCAN, clinical data, cross-validation, YadroSeg, reproducibility, density-based clustering

The new paper is the third and last part of what the author describes as a research triptych. The first two installments dealt with box score data and play-by-play data respectively, each feeding into a growing analytical infrastructure for Spanish basketball. Box scores capture cumulative statistics such as points, rebounds, and assists, while play-by-play records describe the sequence of events during a game. Shooting data, however, adds a crucial spatial dimension. For each attempt, the dataset records the location of the shot on the court, the shot type, and the outcome. This is precisely the kind of information that has transformed how analysts, coaches, and fans understand the modern game, from the rise of the three-point shot to the strategic value of shots near the rim.

To build the dataset, Vinué developed code in the R programming language that automatically collects shooting information from the official ACB website using a web scraping procedure. Web scraping involves programmatically extracting data from web pages, and the author designed the procedure to be friendly and reproducible. The approach draws on established R tools for harvesting web content, including packages such as rvest and RSelenium, and respects the target website’s robots.txt file, which specifies rules for automated access. Once collected, the raw data required ordinary cleaning to be structured into a standard format suitable for analysis. The author then validated the accuracy of the scraped data by plotting the shots and comparing their coordinates and outcomes against the shot charts displayed on the source website, ensuring that the automated pipeline faithfully reproduces the original information.

The scale of the resulting dataset is substantial. It covers the 2024-2025 ACB season and includes 39,550 shots drawn from 306 games and 293 players. Each record contains, among other details, the location of the attempted shot, the shot type, and whether the attempt was successful. A dataset of this size allows for detailed shooting profiles of individual players, showing where on the floor they prefer to shoot and how effective they are from each area. Such profiles go far beyond a simple field goal percentage, which collapses all attempts into a single number regardless of where they were taken. Spatially aware statistics have become central to basketball analytics worldwide, and this work brings that capability to Spanish league data in a systematic, reproducible way.

A key contribution of the study is the integration of shooting data into a previously developed data platform, so that all analyses are centralized in a single place. The platform, described in an earlier paper, already handled box score and play-by-play information; adding shooting data completes the triptych and turns the platform into a comprehensive resource for Spanish professional basketball. Alongside the platform, the author updated BAwiR, an accompanying R package that is freely available from CRAN, the comprehensive archive network for R software. The package includes documentation for all its files as well as three vignettes that guide users through its functionality. Making both the data and the code openly available means that other researchers, analysts, and enthusiasts can reproduce the results and extend the analysis to their own questions.

To make the data accessible to non-programmers, the study also provides a web application for easy interaction. Built with Shiny, a popular framework for turning R code into interactive web tools, the application lets users explore shooting data without writing any code themselves. Interactive visualization has become an essential bridge between raw data and decision-makers in sport, allowing coaches and analysts to query specific players, games, or shot types on demand. The author also gathered feedback from users through a questionnaire, reflecting a growing recognition in the visualization research community that sports tools should be designed in consultation with the people who will actually use them.

One of the more technically interesting parts of the work concerns how to divide the basketball court into meaningful zones. Classical zones such as the paint, the mid-range areas, and the corners are usually defined manually, but the paper explores whether automatic clustering algorithms can do the job instead. Three methods were tested on the shooting dataset: DBSCAN, HDBSCAN, and k-means. DBSCAN, which stands for Density-Based Spatial Clustering of Applications with Noise, groups points that lie close together in space and marks isolated points as noise, but it requires the user to specify a neighborhood radius and a minimum number of points in advance. HDBSCAN extends this idea by building a hierarchy of clusters of varying density, reducing the need to tune the radius parameter. K-means, the simplest of the three, iteratively assigns each shot to its nearest cluster center and then recalculates those centers as the average of the assigned points, with the number of clusters chosen beforehand, often through silhouette analysis.

The results of the clustering comparison were instructive. K-means produced the closest representation of the classical basketball regions, identifying six zones, although it did not distinguish between two-point and three-point shots, and some regions contained both. HDBSCAN roughly identified some areas, such as the paint and the left and right three-point zones, but its other regions were messy. DBSCAN fared worst, returning just two large regions that carry little basketball meaning. When the analysis was refined by filtering first by shot type, k-means divided the two-point area into three regions approximating the paint and the left and right mid-ranges, while the three-point area split into six clearly identifiable regions: left and right corners, left and right elbows, center three-point shots, and long-distance three-point attempts. These regions largely coincide with the manual partition proposed by the author, and the clustering even suggested numerical limits for the court zones, such as paint boundaries at y-coordinates of around plus or minus 2000. The conclusion is that k-means can serve as a useful starting point for determining the coordinates needed to build court regions manually.

The study is candid about the limitations of automated data collection. The scraping code is reproducible with similar shot data, but if the target website changes its structure, manual updates and further cleaning may be required. This is a familiar trade-off in sports analytics, where researchers often depend on websites they do not control. The author acknowledges the official ACB website for making the data accessible, and the paper’s appendices provide a step-by-step guide to the data collection strategy, along with a glossary of basketball terms to help readers less familiar with the sport’s vocabulary. Such transparency supports reproducibility, a value increasingly emphasized in the R sports analytics community, where packages for basketball, football, and other sports are systematically reviewed and shared.

The broader significance of the work lies in democratizing access to professional basketball data outside the NBA. While the NBA offers well-documented application programming interfaces that have spawned a rich ecosystem of analytical tools, European leagues have historically lagged in data availability. Research has shown that analytics investment can influence team performance in professional basketball, and systematic reviews have highlighted both the potential and the uneven adoption of analytical techniques in the sport. By centralizing box score, play-by-play, and shooting data for the ACB in one platform, backed by a free R package and an interactive web application, this research gives Spanish basketball a data infrastructure comparable in spirit to what NBA analysts enjoy. From shot charts and spatial clustering to player shooting profiles, the tools are now in place for coaches, journalists, and fans to interrogate the game at a level of detail that was previously out of reach for Spain’s top league.

Subject of Research: Web scraping and visualization of shooting data from the Spanish ACB basketball league

Article Title: Adding shooting data to a centralized platform for basketball data visualization

Article References: Adding shooting data to a centralized platform for basketball data visualization. (n.d.). https://doi.org/10.1007/s11042-026-21930-2

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21930-2

Keywords: basketball analytics, ACB, shooting data, web scraping, R package, data visualization, Shiny web application, k-means clustering, DBSCAN, HDBSCAN, shot charts, sport analytics