Position-specific lower extremity performance and machine learning-based injury-risk screening classification in adult male basketball players: a random forest and SHAP approach

Abstract

Background

Basketball imposes high position-dependent neuromuscular demands that predispose athletes to lower extremity injuries. Machine learning offers the potential to identify multivariable risk-screening patterns that single-threshold approaches miss. This study aimed to (1) compare position-specific lower extremity performance profiles across five playing positions and (2) develop and explain a Random Forest (RF) classifier of a composite, evidence-based injury-risk screening index in adult male basketball players, interpreted via SHapley Additive exPlanations (SHAP). No injuries were prospectively observed; the outcome is therefore a screening classification and not an observed injury endpoint.

Methods

One hundred adult competitive male basketball players (age 24.5 ± 3.8 years, range 18–30; 20 per position) completed a single-session test battery. Standing height and body mass were measured under standardised conditions immediately before testing, and body mass index (BMI) was computed as body mass (kg) divided by height squared (m²). Performance was assessed with the DeepSport AI-based motion analysis application: countermovement jump height and peak power (CMJ), reactive strength index (RSI), 10-m sprint time, 505 agility time, bilateral single-leg hop distance, and limb asymmetry. Sample size was determined using G*Power 3.1.9.7 (Cohen’s f = 0.45, α = 0.05, 1 − β = 0.95) [19]. A multi-factor threshold model defined the binary screening label. Athletes meeting ≥ 2 of five predefined criteria (limb asymmetry > 15%, RSI < 1.8, 10-m sprint > 1.95 s, 505 agility > 5.0 s, BMI > 27 kg/m²) were classified as high-risk; these five variables also appear among the 13 candidate predictors. An RF classifier was tuned by grid search under stratified 5-fold cross-validation with all preprocessing fitted inside each training fold, and benchmarked against logistic regression, a support vector machine and gradient boosting. SHAP values were computed with TreeSHAP on the complete sample (n = 100); out-of-fold permutation importance was used as the primary importance measure. A pre-specified sensitivity analysis re-fitted the model after removing all five label-defining variables.

Results

Sixty athletes (60%) were classified as high-risk and 40 (40%) as low-risk. BMI differed significantly across positions (F = 3.54, p = 0.010, η² = 0.130) as did RSI (F = 2.50, p = 0.048, η² = 0.095). The RF classifier achieved AUC = 0.947, accuracy = 84.0%, sensitivity = 93.3%, specificity = 70.0%, F1 = 87.5% and κ = 0.655 (mean across 10 cross-validation seeds: AUC = 0.930 ± 0.012). Logistic regression performed comparably (AUC = 0.910, accuracy = 87.0%, κ = 0.726), indicating that the additional complexity of the RF yielded only a marginal discrimination gain. When the five label-defining variables were removed, discrimination fell to AUC = 0.683 (accuracy 62.0%, κ = 0.167), quantifying the extent to which the headline performance reflects recovery of the labelling rule rather than independent signal. Permutation importance ranked limb asymmetry first (ΔAUC = 0.193 ± 0.043), followed by BMI (0.094 ± 0.057) and 10-m sprint time (0.092 ± 0.008).

Conclusions

A Random Forest model incorporating field-based DeepSport performance metrics reproduces a composite lower-extremity risk-screening classification with outstanding internal discrimination, but its predictive validity for actual injury remains untested and it should be regarded as a screening-support tool rather than an injury-prediction model. Limb asymmetry, BMI and 10-m sprint time are the dominant screening factors, although this ordering is partly determined by the labelling rule, and BMI should be read as a proxy for body size and joint loading rather than as a modifiable index of adiposity. SHAP explanations provide practitioner-readable rationales for individual risk scores, supporting integration into athlete monitoring workflows.

Trial registration

ClinicalTrials.gov, NCT07677371; registered 18 June 2026, i.e. retrospectively, during the May–June 2026 data-collection period.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s13102-026-02033-9.

Introduction

Basketball is a high-intensity intermittent sport characterised by repeated explosive actions including vertical jumps, rapid accelerations, decelerations, lateral cuts, and single-leg landings [1]. These demands impose substantial neuromuscular stress on the lower extremity, making musculoskeletal injuries—particularly of the ankle, knee, and hip—among the most prevalent sources of athlete unavailability at all competitive levels [2, 3].

Although neuromuscular control and muscle–tendon adaptation continue to develop into early adulthood [4], the present cohort comprised skeletally mature adult players (age range 18–30 years); maturation-related explanations are therefore not invoked in interpreting the present findings.

The playing position is a fundamental organisational unit in basketball and has well-established links to physical demands: guards (Point Guard, Shooting Guard) emphasise speed, agility, and change-of-direction capacity; forwards (Small and Power Forward) combine agility with explosive power; centres focus on maximal vertical force production, body mass, and contact stability [5, 6]. Position-specific injury profiles and risk thresholds may therefore differ meaningfully, yet most screening protocols apply population-level cut-offs that ignore this variation [7–9].

Traditional injury risk screening relies on a limited number of single-variable thresholds—most commonly limb symmetry index (LSI) or isolated jump height—to classify athletes as at-risk [7–9]. While practical, such approaches fail to capture the multivariate, interacting nature of injury aetiology. Machine learning (ML) algorithms—particularly ensemble methods such as Random Forests—are well-suited to model these interactions, handle non-linear relationships, and produce probabilistic outputs that support risk stratification [10–12].

It must be stated at the outset that screening tests, whether univariate or multivariate, have repeatedly failed to predict individual injury occurrence prospectively [13]. The present work therefore does not claim injury prediction. Its outcome is a composite, evidence-based screening classification, and the modelling exercise asks a narrower question: which field-measurable characteristics carry the most information about that classification, and can this information be expressed transparently enough to support practitioner decision-making?

A practical barrier to clinical adoption of ML models has been their “black-box” character. SHapley Additive exPlanations (SHAP), grounded in cooperative game theory, address this by assigning each feature a marginal contribution to each individual prediction, enabling both global feature-importance ranking and local (per-athlete) explanation [11, 14]. SHAP has been applied to injury risk in team sports [10, 12, 15] but remains underexplored in basketball.

Accurate and accessible performance assessment is a prerequisite for reliable risk modelling. The DeepSport application is an AI-based motion analysis system that uses standard camera devices (laptops, smartphones, tablets) without proprietary hardware [16]. Its validity and reliability for countermovement jump and vertical jump assessment have been confirmed (ICC > 0.90) in comparison with the OptoJump optical system, with no significant systematic bias [16]. This positions DeepSport as a cost-effective field tool suitable for athlete screening.

The objectives of this study were twofold: (1) to characterise and compare lower extremity performance profiles across five playing positions in adult male basketball players using DeepSport, and (2) to develop a Random Forest screening-index classifier interpreted via SHAP, thereby identifying the most influential correlates of elevated screening risk. We hypothesised that (a) significant between-position differences would emerge for at least one neuromuscular variable, and (b) the RF model would achieve excellent discrimination (AUC ≥ 0.85) while SHAP analysis would confirm limb asymmetry and reactive strength as primary predictors.

Methods

Study design and participants

A cross-sectional, observational comparative design was employed. One hundred adult competitive (amateur-league) male basketball players participated. Participants were stratified by primary playing position as assigned by their head coach: Point Guard (PG, n = 20), Shooting Guard (SG, n = 20), Small Forward (SF, n = 20), Power Forward (PF, n = 20), and Center (C, n = 20). The cohort mean age was 24.5 ± 3.8 years (range 18–30 years). All testing was conducted at Sintech Prosthetics and Orthotics Health and Technology Company (Sintech, Turkey) between May and June 2026 under standardised environmental conditions (controlled lighting, surface, and temperature).

Reporting follows the STROBE statement for cross-sectional observational studies [17] and, for the prediction-modelling component, the TRIPOD statement for the transparent reporting of multivariable prediction models [18]. Completed STROBE and TRIPOD checklists are provided as additional files.

In addition, the present study adheres to the CONSORT (Consolidated Standards of Reporting Trials) guidelines. Participant eligibility and enrolment, the flow of participants through the study, the a priori definition and blinded assessment of the outcome, the handling of missing data, and the statistical reporting were documented in accordance with the CONSORT recommendations, which were applied to the present non-randomised design; a completed CONSORT checklist is available from the corresponding author on request.

Inclusion criteria were: (1) male sex, (2) licensed basketball player, (3) ≥ 1 year of continuous basketball training, (4) ≥ 3 training sessions per week, and (5) free of acute lower extremity injury in the 6 months preceding testing. Exclusion criteria were: (1) lower extremity surgery within the preceding 6 months, (2) acute musculoskeletal injury on the test day, (3) neurological, vestibular, or orthopaedic condition affecting balance or motor performance, and (4) incomplete test battery due to any cause.

Sample size calculation

Sample size was determined a priori using G*Power software (version 3.1.9.7; Faul et al. [19]) for the primary aim (between-position comparison). A one-way ANOVA model with five groups was specified; effect size Cohen’s f = 0.45 (large [20]), α = 0.05, 1 − β = 0.95, yielding a minimum required sample of 80 participants. To ensure balanced allocation and allow for data loss the target was increased to n = 100 (20 per position). This calculation powers the group-comparison aim only. For the classification aim, the 60 positive cases and 13 candidate predictors correspond to 4.6 events per variable (EPV), below the conventional minimum of 10 [21] and below the sample size that would be required by contemporary criteria for developing a stable multivariable prediction model [22]. The classification analyses are therefore explicitly framed as exploratory and internally validated only; the consequences of this constraint are quantified in the stability analyses reported below.

Measurement protocol and instrumentation

All performance assessments were conducted using the DeepSport AI-based motion analysis application (DeepSport, Turkey) [16]. DeepSport employs pose estimation and temporal movement analysis through neural networks, enabling kinematic assessment from standard camera devices without proprietary hardware [16]. Its validity and reliability for CMJ and vertical jump assessment have been confirmed (ICC > 0.90; CV 2.12–4.95%; no systematic Bland–Altman bias) against the OptoJump optical reference system [16]. All measurements were captured under standardised camera angles and distances specified by the manufacturer [16].

All assessments were completed in a single session following a standardised 10–15 min warm-up (5 min light jog, dynamic stretching, submaximal jump and agility preparation). Test order was fixed and identical for every participant and progressed from the least to the most neuromuscularly demanding task: (1) CMJ, (2) RSI, (3) single-leg hop, (4) 10-m sprint, (5) 505 agility. This sequence follows the graded functional-testing hierarchy set out by Reiman and Manske [23], in which bilateral, vertically loaded, self-paced tasks are administered before unilateral tasks, and unilateral tasks before maximal horizontal acceleration and change-of-direction tasks. The rationale is that eccentric and deceleration loading, and therefore the rate at which fatigue accumulates, increases across that progression: the 505 agility test combines maximal acceleration with a 180° deceleration–re-acceleration cycle and is the most demanding item in the battery, whereas the CMJ is submaximal in its horizontal demand and largely self-paced. Administering the battery in this order limits carry-over of fatigue from high-demand to low-demand items and avoids the systematic underestimation of jump height and reactive strength that occurs when maximal change-of-direction work precedes them [23]. Rest intervals were identical for all participants: exactly 2 min between successive trials of the same test and exactly 3 min between different tests, timed with a digital stopwatch and verified against the session record sheets. Participants remained standing or seated at rest during these intervals and performed no additional activity. Recovery duration was therefore uniform across participants and could not differentially advantage any individual or position group.

Anthropometry and body mass index

Anthropometric measures were taken before the warm-up, in light sportswear and without shoes, by the same trained assessor for all participants. Standing height was measured to the nearest 0.1 cm with a stadiometer, with the participant’s head aligned in the Frankfort horizontal plane at the end of a deep inspiration, and body mass to the nearest 0.1 kg on a calibrated digital scale. Body Mass Index (BMI) was then calculated as body mass (kg) divided by height squared (m²). BMI is used in two distinct roles in this manuscript and both are stated explicitly: it is a descriptive anthropometric variable compared between playing positions (Table 1), and it is one of the five criteria that define the composite screening label (BMI > 27 kg/m²).

Table 1.

Anthropometric and background characteristics by playing position (Mean ± SD)

Variable	PG (n = 20)	SG (n = 20)	SF (n = 20)	PF (n = 20)	C (n = 20)	F	p	η²
Age (yr)	25.4 ± 3.9	24.8 ± 3.6	23.9 ± 4.2	24.7 ± 4.1	24.0 ± 3.2	0.55	0.702	0.023
Height (cm)	195.2 ± 11.4	191.0 ± 11.2	194.9 ± 11.5	198.5 ± 8.8	191.5 ± 10.6	1.60	0.181	0.063
Weight (kg)	100.2 ± 11.7	99.9 ± 15.0	97.0 ± 15.9	88.7 ± 15.4	95.1 ± 12.4	2.18	0.077	0.084
BMI (kg/m²)	26.7 ± 5.0	27.7 ± 5.4	25.8 ± 5.4	22.5 ± 3.4	26.0 ± 3.7	3.54	0.010**	0.130
Sport Exp. (yr)	9.5 ± 3.5	10.1 ± 3.2	9.0 ± 3.2	10.5 ± 3.3	9.6 ± 3.5	0.56	0.695	0.023

PG  Point Guard, SG  Shooting Guard, SF  Small Forward, PF  Power Forward, C  Center

F = one-way ANOVA F-statistic, df = (4, 95); p-values from the exact F distribution

** p < 0.01. η² bands: small 0.01 to < 0.06, medium 0.06 to < 0.14, large ≥ 0.14

Countermovement Jump (CMJ)

Participants performed three maximal bilateral CMJs with hands on hips, following a rapid countermovement to a self-selected depth. Jump height (cm) and peak power (W) were derived from flight-time analysis by the DeepSport system [16]. The best of three trials was retained. This protocol is consistent with the validated assessment conditions reported by Aydemir et al. [16].

Reactive Strength Index (RSI)

RSI was assessed via drop-jump from a standardised box height. Participants were instructed to minimise ground contact time while maximising jump height. RSI = jump height (m) / ground contact time (s). Flight time and contact time were recorded by DeepSport [16]. The best of three trials was used. RSI reflects stretch-shortening cycle (SSC) efficiency and has been proposed as a sensitive neuromuscular risk indicator [7].

Single-leg hop test

Horizontal single-leg hop distance was assessed bilaterally using the DeepSport application [16]. Participants performed a maximal single-leg forward hop and were required to stabilise on the landing limb. Three valid trials were completed per limb (right and left); the best was retained. Limb asymmetry was calculated as: . Values exceeding 15% have been associated with elevated injury risk [8, 24].

10-m sprint test

Acceleration capacity was assessed over 10 m from a standing start, with the front foot placed 0.5 m behind the starting line. Timing was initiated by the DeepSport pose-estimation algorithm at first movement and stopped at the 10-m mark [16]. The better of two trials (2 min recovery between trials, in line with the fixed inter-trial interval specified above) was used. Sprint time is considered a proximal indicator of lower limb neuromuscular power and explosive capacity [25].

505 agility test

Change-of-direction (COD) ability was assessed using the 505 agility test, originally described by Draper and Lancaster [26], which requires a 180° turn at the 5-m mark of a 10-m course. Total completion time (s) was recorded via DeepSport [16]. Participants started from a standing position with the front foot placed 0.5 m behind the starting line, as in the 10-m sprint test. The dominant-leg pivot trial was used. COD performance is position-sensitive in basketball and is associated with injury risk through asymmetric deceleration loading [27].

Measurement reliability

Device-level validity and reliability evidence for DeepSport is currently restricted to CMJ and vertical jump assessment (ICC > 0.90; CV 2.12–4.95% against OptoJump) [16]. Applied to the present sample CMJ mean of 44.8 cm, this CV range corresponds to a typical error of measurement of approximately 0.95–2.22 cm. No published device-specific validity or reliability coefficients exist for the RSI, 10-m sprint, single-leg hop or 505 agility modules of the application. Because only the best-of-trial value was retained in the analysis dataset, trial-to-trial within-study reliability (ICC, coefficient of variation, standard error of measurement) could not be recomputed retrospectively for these four tests. This is an acknowledged constraint on the measurement basis of the composite screening index and is restated in the Limitations; prospective work by this group will retain all trials so that within-study reliability can be reported directly.

Injury-risk screening classification

No injuries were recorded during or after the testing session. The outcome analysed here is therefore not an observed injury but a composite, evidence-based screening index constructed from the measured variables themselves. Athletes accumulating ≥ 2 of the following five risk factors were classified as high-risk (binary label: 1 = high risk, 0 = low risk):

• Limb asymmetry > 15%: impaired bilateral load distribution [8, 24].

• RSI < 1.8: reduced reactive neuromuscular control [7].

• 10-m sprint time > 1.95 s: below-average acceleration capacity [25].

• 505 agility time > 5.0 s: impaired change-of-direction mechanics [27].

• BMI > 27 kg/m²: elevated relative body mass loading [28].

Rationale for the “≥ 2 of 5” rule

The scheme follows the multifactorial model of injury aetiology [9], in which risk emerges from the accumulation of several sub-optimal characteristics rather than from any single deficit, and mirrors the cumulative-criterion logic used in established multi-item field screening batteries [24, 29]. A single criterion was rejected as over-sensitive: in this sample, 89 of 100 athletes met at least one criterion (Table 3), which would have produced an uninformative 89% “high-risk” prevalence. A threshold of ≥ 3 was rejected as over-specific (30% prevalence) and would have left too few negative cases for stable model evaluation. The ≥ 2 rule yielded a 60/40 split. We emphasise that these cut-offs are conventions imported from the literature and not empirically derived in this cohort; the sensitivity of the whole analysis to this choice is examined below.

Table 3.

Frequency of individual screening criteria and distribution of criterion counts (n = 100)

Screening criterion	n meeting criterion	%
Limb asymmetry > 15%	41	41.0
RSI < 1.8	26	26.0
10-m sprint > 1.95 s	27	27.0
505 agility > 5.0 s	52	52.0
BMI > 27 kg/m²	37	37.0
Criteria met: 0	11	11.0
Criteria met: 1	29	29.0
Criteria met: 2	30	30.0
Criteria met: 3	26	26.0
Criteria met: 4	4	4.0
Criteria met: 5	0	0.0

A threshold of ≥ 1 criterion would classify 89% of the cohort as high-risk; ≥ 3 would classify 30%. The adopted ≥ 2 rule yields 60%

Relationship between the label and the model inputs

All five variables that define the label (limb asymmetry, RSI, 10-m sprint time, 505 agility time and BMI) also appear among the 13 candidate predictors. By construction, therefore, part of the classifier’s discrimination reflects the model recovering the deterministic rule that generated the labels, and any SHAP finding that risk rises near a pre-set cut-off partly re-expresses that cut-off rather than discovering it independently. This is stated explicitly here, rather than only in the Limitations, and is quantified by the pre-specified sensitivity analysis described in the next section.

Random forest classifier and benchmarking

A Random Forest classifier [30] was used as the primary model. All modelling was performed in Python 3.10.12 with scikit-learn 1.5.2 [31], NumPy 2.2.6 and pandas 2.3.3. The reference scikit-learn implementation was used so that the analysis is fully reproducible; the global random seed was fixed at 42 and is reported in the analysis script provided as an additional file.

Every preprocessing step was fitted inside the training partition only. Specifically, standardisation (µ = 0, σ = 1) was implemented as the first stage of a scikit-learn Pipeline, so that the scaler is estimated on each training fold and applied to the corresponding held-out fold. No transformation was fitted on the complete dataset before splitting, and no information leakage from the held-out data into model fitting occurred.

Hyperparameters were selected by exhaustive grid search (72 candidate configurations) maximising cross-validated AUC: n_estimators ∈ {100, 300, 500}, max_depth ∈ {3, 4, 6, None}, min_samples_split ∈ {2, 5, 10}, min_samples_leaf ∈ {1, 2}, max_features = √p. The selected configuration was n_estimators = 300, max_depth = None, min_samples_split = 2, min_samples_leaf = 2, max_features = √p (p = 13).

Three benchmark models were evaluated alongside the Random Forest: unregularised logistic regression (penalty = none, lbfgs solver, max_iter = 1000; the conventional statistical baseline), a support vector machine with an RBF kernel, and gradient boosting. All four models were evaluated under the identical pipeline, cross-validation scheme and seed. The benchmark models were run at the scikit-learn 1.5.2 default settings (support vector machine: C = 1.0, γ = “scale”, probability estimates enabled; gradient boosting: n_estimators = 100, learning_rate = 0.1, max_depth = 3) and were deliberately left untuned, so that they represent the performance a practitioner would obtain without a hyperparameter search.

Model performance was evaluated under stratified 5-fold cross-validation; out-of-fold predicted probabilities were concatenated across folds before computing metrics. Because a single fold assignment can markedly affect estimates at n = 100, the entire cross-validation was repeated across 10 independent seeds and results are reported as mean ± SD across seeds in addition to the seed-42 point estimates.

Pre-specified sensitivity analyses

Two additional models were fitted. (i) Non-defining model: all five label-defining variables were removed, leaving eight predictors (age, height, weight, sport experience, CMJ height, CMJ peak power, right hop, left hop) that do not enter the labelling rule. The discrimination of this model indexes the signal available from variables that are not, by construction, tied to the outcome. (ii) Collinearity-pruned model: weight, height and the two raw hop distances were removed because BMI is an algebraic function of the first two and asymmetry of the last two, leaving nine non-redundant predictors.

Feature importance and SHAP analysis

Mean decrease in Gini impurity is known to be biased towards continuous and high-cardinality predictors [32]. It is therefore reported here only as a secondary measure. The primary importance measure is out-of-fold permutation importance, computed as the mean decrease in held-out AUC over 30 random permutations of each feature within each of the five cross-validation test folds (mean ± SD across folds).

SHAP values were computed with the exact TreeSHAP algorithm [14] on the complete sample (n = 100). TreeSHAP is polynomial-time for tree ensembles, so no subsampling was required.

Statistical analysis

Data are presented as mean ± standard deviation (SD). The D’Agostino–Pearson omnibus test (K²) [33] was used to evaluate normality within each position group. Between-position differences were tested with one-way ANOVA. p-values were obtained from the exact F distribution with df = (4, 95). Effect sizes are reported as η² with the following interpretation bands [20]: small 0.01 to < 0.06, medium 0.06 to < 0.14, large ≥ 0.14. Between-group (high- versus low-risk) comparisons used Welch’s independent-samples t-test, which does not assume equal variances, and Cohen’s d was computed with the mean of the two group variances as the standardiser. Statistical significance was set at α = 0.05. Model discrimination was quantified by the AUC of the ROC curve, interpreted using the conventional bands of Hosmer and Lemeshow [34] (0.7–0.8 acceptable, 0.8–0.9 excellent, > 0.9 outstanding). Agreement beyond chance was quantified by Cohen’s Kappa [35] and interpreted using the descriptive bands of Landis and Koch [36].

Statistical software. All analyses were carried out in Python 3.10.12 on a 64-bit Linux platform; no proprietary statistical package (SPSS, SAS, Stata or equivalent) was used at any stage. Data handling and matrix operations used NumPy 2.2.6 and pandas 2.3.3. Descriptive statistics, the D’Agostino–Pearson K² normality test, one-way analysis of variance and independent-samples t-tests were computed with SciPy 1.14.1 [37]. Eta-squared effect sizes, variance inflation factors and the supporting regression diagnostics were computed with statsmodels 0.14.4 [38]. All machine-learning procedures — pipeline construction with fold-internal standardisation, the hyperparameter grid search, stratified 5-fold cross-validation, the Random Forest, logistic regression, support vector machine and gradient boosting models, and out-of-fold permutation importance — were implemented with scikit-learn 1.5.2 [31]. Shapley values were computed with the shap package 0.46.0 using the exact TreeSHAP algorithm [14]. Figures were produced with Matplotlib 3.9.2 [39] and seaborn 0.13.2. The a priori sample size calculation was performed in G*Power 3.1.9.7 [19]. The global random seed was fixed at 42 for every stochastic procedure, and the complete analysis script, together with the de-identified dataset, is supplied as an additional file so that every value reported in this manuscript can be reproduced exactly.

Ethical approval

This study was approved by the Non-Interventional Clinical Research Ethics Committee of Iğdır University (Decision No.: 27/10, Date: 06 May 2026; Reference No.: E-37077861-900-217860). All procedures were conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants. No participant was younger than 18 years (minimum age 18 years); accordingly, no parental or guardian consent was required.

Results

Participant characteristics

Anthropometric and background characteristics are presented in Table 1. Groups did not differ significantly in age (F = 0.55, p = 0.702), height (F = 1.60, p = 0.181), body mass (F = 2.18, p = 0.077), or sport experience (F = 0.56, p = 0.695). A statistically significant between-position difference was identified for BMI (F = 3.54, p = 0.010, η² = 0.130, medium effect, at the upper boundary of the medium band), with SG recording the highest mean (27.7 ± 5.4 kg/m²) and PF the lowest (22.5 ± 3.4 kg/m²) [28]. The cohort age range was 18–30 years, confirming a skeletally mature adult sample.

Position-specific lower extremity performance

Performance data by position are presented in Table 2. Of the eight performance variables, only RSI reached statistical significance (F = 2.50, p = 0.048, η² = 0.095, medium effect). Centers recorded the highest mean RSI (2.68 ± 0.70) and Shooting Guards the lowest (2.16 ± 0.71), consistent with position-specific demands for reactive landing control [7]. CMJ height (F = 1.13, p = 0.349), 10-m sprint (F = 0.67, p = 0.617), 505 agility (F = 2.11, p = 0.086), single-leg hop distances, and limb asymmetry did not differ significantly between positions (all p > 0.05). Normalised position profiles are shown in Fig. 1 and variable distributions in Fig. 2.

Table 2.

Lower extremity performance variables by playing position (Mean ± SD)

Variable	PG	SG	SF	PF	C	F	p	η²
CMJ Height (cm)	46.3 ± 9.4	42.1 ± 8.5	45.7 ± 10.4	47.2 ± 9.0	42.9 ± 9.2	1.13	0.349	0.045
CMJ Peak Power (W)	5138 ± 917	4994 ± 880	5156 ± 911	4963 ± 894	4862 ± 956	0.37	0.831	0.015
RSI	2.49 ± 0.57	2.16 ± 0.71	2.20 ± 0.55	2.24 ± 0.63	2.68 ± 0.70	2.50	0.048*	0.095
Sprint 10 m (s)	1.90 ± 0.12	1.84 ± 0.13	1.83 ± 0.18	1.85 ± 0.16	1.83 ± 0.16	0.67	0.617	0.027
Agility 505 (s)	5.08 ± 0.47	5.18 ± 0.52	4.87 ± 0.48	4.80 ± 0.42	5.02 ± 0.50	2.11	0.086	0.082
Right Hop (m)	2.25 ± 0.24	2.19 ± 0.27	2.32 ± 0.28	2.36 ± 0.32	2.26 ± 0.36	0.93	0.448	0.038
Left Hop (m)	2.36 ± 0.22	2.26 ± 0.26	2.33 ± 0.34	2.27 ± 0.30	2.29 ± 0.35	0.39	0.814	0.016
Asymmetry (%)	11.2 ± 7.9	12.9 ± 7.9	16.3 ± 9.8	12.4 ± 8.5	14.7 ± 9.9	1.04	0.392	0.042

Values are Mean ± SD; all data collected via the DeepSport AI application [16]. F = one-way ANOVA, df = (4, 95), exact p-values. * p < 0.05. η² bands as in Table 1

CMJ countermovement jump, RSI  reactive strength index

Fig. 1.

Position-specific performance radar chart displaying normalised (0–1) mean profiles across six variables. Sprint and agility times were inverted so that higher values represent better performance. PG = Point Guard; SG = Shooting Guard; SF = Small Forward; PF = Power Forward; C = Center

Fig. 2.

Distribution of key performance variables across playing positions (DeepSport assessment [16]). Boxes: interquartile range; horizontal lines: medians; whiskers: 1.5 × IQR. * p < 0.05; ** p < 0.01; ns = not significant

Composite screening classification

Applying the multi-factor threshold model, 60 athletes (60%) were classified as high-risk and 40 (40%) as low-risk. Risk prevalence by position was: SG 70% (n = 14), C 65% (n = 13), PG and SF 60% each (n = 12), and PF 45% (n = 9). The frequency with which each individual criterion was met, and the resulting distribution of criterion counts, are given in Table 3; these figures underpin the justification for the ≥ 2 rule set out in the Methods.

High-risk athletes exhibited higher mean limb asymmetry (16.4 ± 9.0%) than low-risk athletes (9.0 ± 6.5%; t = 4.78, p < 0.001, Cohen’s d = 0.94). Corresponding group differences were observed for BMI (27.5 ± 4.7 vs. 23.0 ± 3.7 kg/m², d = 1.05), 10-m sprint (1.88 ± 0.15 vs. 1.80 ± 0.14 s, d = 0.54), 505 agility (5.09 ± 0.49 vs. 4.84 ± 0.46 s, d = 0.52) and RSI (2.23 ± 0.68 vs. 2.54 ± 0.56, d = − 0.50). Figure 3 illustrates the classification by position and the asymmetry distribution by group.

Fig. 3.

a Stacked bar chart of screening classification by playing position; percentages denote the proportion classified as high-risk. b Violin plot of limb asymmetry (%) by screening classification; dashed line = 15% clinical threshold [8, 24]

Classifier performance and model benchmarking

Under stratified 5-fold cross-validation with all preprocessing fitted within training folds, the tuned Random Forest achieved AUC = 0.947, accuracy = 84.0%, sensitivity = 93.3%, specificity = 70.0% and κ = 0.655 (Tables 4 and 5; the corresponding confusion matrix is shown in Fig. 4). Across 10 independent cross-validation seeds the corresponding means were AUC = 0.930 ± 0.012, accuracy = 83.8 ± 1.8%, sensitivity = 91.8 ± 1.2%, specificity = 71.8 ± 4.1% and κ = 0.653 ± 0.039, indicating that the point estimates are not an artefact of a single favourable fold assignment.

Table 4.

Benchmarking of classification algorithms (stratified 5-fold cross-validation, seed = 42, n = 100)

Model	AUC	Accuracy	Sens.	Spec.	PPV	NPV	F1	κ
Random Forest (tuned, 13 features) — primary	0.947	84.0%	93.3%	70.0%	82.4%	87.5%	87.5%	0.655
Random Forest (pre-specified hyperparameters)	0.932	82.0%	90.0%	70.0%	81.8%	82.4%	85.7%	0.615
Random Forest (collinearity-pruned, 9 features)	0.947	85.0%	91.7%	75.0%	84.6%	85.7%	88.0%	0.681
Gradient Boosting	0.938	82.0%	81.7%	82.5%	87.5%	75.0%	84.5%	0.631
Logistic Regression (baseline)	0.910	87.0%	91.7%	80.0%	87.3%	86.5%	89.4%	0.726
Support Vector Machine (RBF)	0.882	82.0%	91.7%	67.5%	80.9%	84.4%	85.9%	0.612

All models were evaluated under an identical pipeline (fold-internal standardisation), cross-validation scheme and random seed

Sens. Sensitivity, Spec. specificity, PPV/NPV  positive/negative predictive value, κ  Cohen’s Kappa

Table 5.

Primary random forest performance metrics (5-fold cross-validation, n = 100)

Metric	Value (seed 42)	Mean ± SD (10 seeds)	Interpretation
Accuracy	84.0%	83.8 ± 1.8%	84 of 100 athletes correctly classified (56 TP + 28 TN)
Sensitivity (Recall)	93.3%	91.8 ± 1.2%	56 of 60 high-risk athletes correctly identified
Specificity	70.0%	71.8 ± 4.1%	28 of 40 low-risk athletes correctly identified
Positive Predictive Value	82.4%	83.0 ± 2.0%	56 of 68 predicted high-risk are true positives
Negative Predictive Value	87.5%	85.4 ± 1.9%	28 of 32 predicted low-risk are true negatives
F1-Score	87.5%	87.2 ± 1.3%	Harmonic mean of sensitivity and PPV
AUC-ROC	0.947	0.930 ± 0.012	Outstanding discrimination of the screening label [34]
Cohen’s Kappa	0.655	0.653 ± 0.039	Substantial agreement beyond chance [36]

Confusion-matrix cell counts reconcile exactly with every derived metric. Ten-seed means quantify the stability of the estimates at this sample size. AUC bands from Hosmer and Lemeshow [34]; Kappa bands from Landis and Koch [36]: 0.00–0.20 slight; 0.21–0.40 fair; 0.41–0.60 moderate; 0.61–0.80 substantial; 0.81–1.00 almost perfect

Fig. 4.

Confusion matrix for the primary Random Forest classifier under stratified 5-fold cross-validation (seed = 42). Rows = actual class; columns = predicted class

Benchmarking (Table 4; Fig. 5a) showed that the advantage of the Random Forest over conventional modelling is modest and not uniform across metrics. Logistic regression achieved a lower AUC (0.910 vs. 0.947) but higher accuracy (87.0% vs. 84.0%), higher specificity (80.0% vs. 70.0%) and higher agreement (κ = 0.726 vs. 0.655). Gradient boosting gave an intermediate AUC (0.938) with the most balanced sensitivity–specificity profile (81.7%/82.5%), and the support vector machine performed worst (AUC = 0.882). Given a difference in AUC of 0.037 between the RF and a model with 13 interpretable coefficients, the additional complexity of the ensemble is justified primarily by its ranking performance and its compatibility with SHAP-based individual explanation, not by a decisive gain in classification accuracy.

Fig. 5.

a Receiver operating characteristic curves for the four candidate algorithms using all 13 features. b Circularity sensitivity analysis: discrimination of the Random Forest with all 13 features versus the same model after removal of the five label-defining variables

Removing the four collinear anthropometric/derived predictors (weight, height, both raw hop distances) left nine non-redundant features and did not degrade performance (AUC = 0.947, accuracy = 85.0%, specificity = 75.0%, κ = 0.681), confirming that the redundant features contribute little unique information.

Sensitivity analysis: dependence of performance on the label-defining variables

When all five label-defining variables were removed, leaving eight predictors that play no part in constructing the outcome, discrimination fell from AUC = 0.947 to AUC = 0.683 (10-seed mean 0.694 ± 0.033), accuracy from 84.0% to 62.0%, specificity from 70.0% to 37.5% and agreement from κ = 0.655 to κ = 0.167 (Table 6; Fig. 5b). The same pattern held for logistic regression (AUC 0.910 → 0.698).

Table 6.

Circularity Sensitivity Analysis: Performance With and Without the Label-Defining Variables

Model / feature set	AUC (seed 42)	AUC (10 seeds)	Accuracy	Sens.	Spec.	κ
Random Forest — all 13 features	0.947	0.930 ± 0.012	84.0%	93.3%	70.0%	0.655
Random Forest — 5 label-defining features removed	0.683	0.694 ± 0.033	62.0%	78.3%	37.5%	0.167
Logistic Regression — all 13 features	0.910	0.891 ± 0.013	87.0%	91.7%	80.0%	0.726
Logistic Regression — 5 label-defining removed	0.698	—	65.0%	73.3%	52.5%	0.262

Removed (label-defining) variables: limb asymmetry, BMI, 10-m sprint time, 505 agility time, RSI. Retained (non-defining) variables: age, height, weight, sport experience. CMJ height, CMJ peak power, right hop distance, left hop distance

The interpretation is unambiguous: approximately 59% of the excess discrimination above chance (ΔAUC = 0.264, i.e. a fall from 0.447 to 0.183 above the chance value of 0.5) is attributable to the model recovering the deterministic labelling rule. The residual AUC of ≈ 0.68 from anthropometry, age, training history, CMJ performance and raw hop distances is modest but non-trivial, and indicates that variables outside the labelling rule do carry some information about the composite screening classification — largely through body size, which correlates with BMI. The headline metrics must not be read as evidence of injury-prediction accuracy.

Feature importance and SHAP analysis

Out-of-fold permutation importance is reported as the primary measure; mean decrease in Gini impurity is reported alongside it as a secondary and known-biased measure [32] (Table 7; Fig. 6). Limb asymmetry was the dominant predictor by every measure (permutation ΔAUC = 0.193 ± 0.043; mean |SHAP| = 0.124), followed by BMI (0.094 ± 0.057; 0.080) and 10-m sprint time (0.092 ± 0.008; 0.065). The three measures agreed closely on ordering (Spearman ρ = 0.94 between permutation importance and mean |SHAP|; ρ = 0.92 between permutation importance and Gini), although Gini and SHAP both ranked body weight — a continuous, high-variance and collinear predictor — several places higher than the permutation measure, which is exactly the bias pattern described by Strobl et al. [32]. The four highest-ranked features accounted for 67.8% of total mean |SHAP|.

Table 7.

Feature importance: out-of-fold permutation importance (primary), Mean |SHAP| and gini (secondary)

Feature	Permutation ΔAUC (mean ± SD)	Rank	Mean |SHAP|	Rank	Gini	Rank
Asymmetry (%)	0.193 ± 0.043	1	0.124	1	0.214	1
BMI (kg/m²)	0.094 ± 0.057	2	0.080	2	0.154	2
Sprint 10 m (s)	0.092 ± 0.008	3	0.065	3	0.121	3
Agility 505 (s)	0.042 ± 0.031	4	0.041	5	0.071	6
RSI	0.035 ± 0.021	5	0.039	6	0.081	5
Weight (kg)	0.028 ± 0.022	6	0.061	4	0.112	4
Height (cm)	0.012 ± 0.015	7	0.026	7	0.060	7
Right Hop (m)	0.010 ± 0.013	8	0.008	11	0.036	9
Left Hop (m)	0.008 ± 0.018	9	0.011	8	0.035	10
Sport Experience (yr)	0.005 ± 0.010	10	0.010	10	0.027	12
CMJ Peak Power (W)	0.004 ± 0.006	11	0.010	9	0.036	8
Age (yr)	0.002 ± 0.010	12	0.005	13	0.021	13
CMJ Height (cm)	0.000 ± 0.004	13	0.007	12	0.031	11

Permutation importance = mean decrease in held-out AUC across 30 permutations within each of five cross-validation test folds (mean ± SD across folds); this is the primary measure. Mean |SHAP| from exact TreeSHAP on the complete sample (n = 100). Gini = mean decrease in impurity, reported as a secondary measure because of its documented bias towards continuous and high-cardinality predictors [32]

Fig. 6.

Feature importance by three complementary measures: (a) out-of-fold permutation importance (primary; error bars = SD across cross-validation folds), (b) mean |SHAP| from exact TreeSHAP on the complete sample, (c) mean decrease in Gini impurity (secondary, biased measure)

The beeswarm plot (Fig. 7) shows the expected directional effects: high limb asymmetry, high BMI, long 10-m sprint times and long 505 agility times all push predictions towards the high-risk class, whereas high RSI pushes predictions towards low risk. Age and sport experience cluster tightly around zero.

Fig. 7.

SHAP beeswarm summary plot computed with exact TreeSHAP on the complete sample (n = 100). Each point represents one athlete–feature pair. Horizontal position = SHAP value; colour = feature value (blue = low, red = high). Features ordered by mean |SHAP|

The dependence plots (Fig. 8) show sharp transitions in SHAP value at approximately 15% asymmetry, 27 kg/m² BMI and 1.95 s sprint time. These inflection points coincide with the cut-offs used to construct the label and are therefore a re-expression of the labelling rule rather than an independent, data-driven rediscovery of biologically valid thresholds.

Fig. 8.

SHAP dependence plots for the three highest-importance predictors. Each point = one athlete; colour = screening classification (green = low risk, red = high risk). Dashed lines = cubic trend fits. Vertical dotted lines mark the a priori labelling thresholds

Discussion

This study characterised position-specific lower extremity performance in adult male basketball players and developed a SHAP-explained Random Forest classifier of a composite, evidence-based screening index derived from DeepSport AI measurements. Four principal findings emerge.

First, between-position performance differences were largely non-significant, with only RSI (η² = 0.095) and BMI (η² = 0.130) reaching statistical significance. This partially contradicts prior work reporting position-specific differences in CMJ and agility at the elite level [5, 6], and may reflect a lower degree of positional training specialisation in amateur-league competition. Because the cohort was skeletally mature (18–30 years), maturation-based explanations are not applicable. The significant BMI difference — SG highest, PF lowest — aligns with body-composition demands [5, 28].

Second, and most consequentially for interpretation, the classifier’s apparent accuracy is substantially an artefact of the way the outcome was constructed. Five of the thirteen predictors are the very variables from which the label is computed, and removing them reduces AUC from 0.947 to 0.683. Any reader encountering the headline figure in isolation would over-estimate what the model can do. We therefore reframe the contribution: the model demonstrates that a compact set of field-measurable variables can reproduce a multi-criterion screening decision with high fidelity and can explain each individual decision transparently — which is operationally useful for automating and auditing a screening workflow — but it does not constitute evidence that the model, or the underlying screening index, predicts injury. Screening tests have a poor record of prospective individual prediction [13], and nothing in the present design can overturn that.

Third, benchmarking against conventional modelling showed only a marginal advantage for the ensemble. Logistic regression achieved AUC = 0.910 against the RF’s 0.947 while exceeding it on accuracy (87.0% vs. 84.0%), specificity (80.0% vs. 70.0%) and Kappa (0.726 vs. 0.655). At n = 100 with 4.6 events per variable, a transparent 13-coefficient model is defensible as the practical default, and the case for the Random Forest rests on its superior ranking performance and its natural pairing with TreeSHAP for per-athlete explanation rather than on superior classification.

Fourth, importance analysis converged on limb asymmetry, BMI and 10-m sprint time as the dominant features across all three importance measures. The ordering is partly determined by the labelling rule, but it is informative that asymmetry dominates so heavily (ΔAUC = 0.193, more than twice the next feature) even though the 505 agility criterion was met more often (52 vs. 41 athletes). Notably, CMJ height and peak power — routinely captured by DeepSport [16] and commonly used as primary screening metrics — ranked last and eleventh respectively by permutation importance. Practitioners relying solely on CMJ height as a risk screen may therefore overlook athletes who are powerful but asymmetric.

Regarding the a priori hypotheses: hypothesis (a) was supported: RSI, a neuromuscular variable, differed significantly between positions (BMI also differed, but is an anthropometric rather than a neuromuscular measure, and so does not by itself satisfy the hypothesis as stated). Hypothesis (b) was only partially supported. The AUC ≥ 0.85 criterion was met and limb asymmetry was confirmed as the leading predictor, but reactive strength was not: RSI ranked fifth by permutation importance and sixth by mean |SHAP|, rather than emerging as a primary predictor as hypothesised.

Interpretation of BMI in this population

BMI ranked second by every importance measure, but its interpretation requires caution in an athletic cohort. BMI does not distinguish fat mass from fat-free mass and systematically misclassifies muscular athletes as overweight [40]; in the present sample the mean BMI was 25.7 kg/m² in men averaging 194 cm and 96 kg, a profile in which elevated BMI frequently reflects lean mass rather than adiposity. BMI should therefore be read here as a proxy for absolute body size and consequent joint loading during landing and deceleration, not as an index of adiposity or of modifiable “excess weight”. Practitioners should not translate a high BMI-driven risk score into a weight-loss recommendation without body-composition assessment.

Feature redundancy

Several predictors are algebraically related: BMI is a deterministic function of height and weight, and limb asymmetry is a deterministic function of the two hop distances. Variance inflation factors confirmed severe collinearity among this subset (BMI 115.6, weight 77.2, height 44.5) while all other predictors were below 1.2, and weight–BMI correlated at r = 0.785. Although Random Forests are relatively robust to collinearity for prediction, importance is diluted and redistributed across correlated features, which is the most likely explanation for weight ranking fourth by Gini and SHAP but only sixth by permutation importance. A collinearity-pruned nine-feature model performed at least as well as the full model (AUC = 0.947, accuracy 85.0%, specificity 75.0%, κ = 0.681) with all VIF < 1.2, and is the configuration we would recommend for any future deployment.

Practical implications of the sensitivity–specificity trade-off

The primary model achieves 93.3% sensitivity at 70.0% specificity. In this cohort that means 12 of 40 low-risk athletes (30%) would be flagged for follow-up unnecessarily, alongside 56 of 60 true high-risk athletes correctly identified. In a 100-athlete programme with the observed 60% prevalence, 68 athletes would be referred to secondary assessment to capture 56 genuine cases — a positive predictive value of 82.4% and roughly 1.2 athletes screened per true case detected. Whether this is acceptable depends on the cost of the downstream response. Where follow-up is a low-cost, low-burden intervention with independent performance benefit — asymmetry-targeted unilateral strength work, landing-mechanics coaching — over-referral is largely benign and the high-sensitivity operating point is appropriate. Where it entails training restriction, imaging, or exclusion from competition, a 30% false-positive rate is not defensible and the decision threshold should be raised: gradient boosting, or the collinearity-pruned RF, offers a more balanced profile (82.5% and 75.0% specificity respectively) at modest sensitivity cost. Practitioners should select the operating point deliberately rather than accept the default 0.5 probability cut-off.

Practical applications

For coaches and certified athletic trainers, three operational messages follow from these data. (1) Single-leg hop asymmetry is the single most informative field measure in this battery and requires only a tape measure and a camera; it should be assessed routinely rather than as an occasional add-on to jump testing. (2) CMJ height alone is a weak screening variable in this context — an athlete can jump well and still present with substantial bilateral imbalance — so jump testing should be supplemented rather than relied upon. (3) The SHAP output gives a per-athlete, feature-level rationale that can be shown directly to the athlete, which supports adherence to corrective programming in a way that an unexplained risk score does not. These recommendations concern the organisation of screening practice; they should not be read as evidence that acting on them will reduce injury incidence, which remains to be tested prospectively.

Limitations

Several limitations warrant acknowledgement, and the first is fundamental. (i) The outcome is a composite screening classification computed from five of the thirteen predictors, not an observed injury. No injuries were recorded, no prospective follow-up was undertaken, and the predictive validity of the model for actual injury is entirely untested. The reported discrimination substantially reflects recovery of the labelling rule, as quantified in Table 6. (ii) With 60 events and 13 candidate predictors the study has 4.6 events per variable, below the conventional minimum of 10 [21] and below contemporary sample-size requirements for prediction-model development [22]. Estimates are correspondingly unstable; the 10-seed cross-validation SDs (AUC ± 0.012, specificity ± 4.1%) bound the fold-assignment component of this instability but not the sampling component, and no external validation was possible. (iii) The five thresholds and the ≥ 2 rule are conventions imported from heterogeneous literatures rather than empirically derived cut-offs, and different defensible choices would produce different labels and hence different models. (iv) Device-specific validity and reliability evidence for DeepSport exists only for CMJ and vertical jump [16]; the RSI, sprint, hop and agility modules that underpin four of the five labelling criteria have no published validation, and because only best-of-trial values were retained, within-study trial-to-trial reliability could not be computed for them. (v) The cohort was exclusively male, drawn from a single geographical region and a single testing facility, limiting generalisability. (vi) Potentially important predictors — training load, sleep, prior injury history, playing exposure — were not collected. Future work should follow a prospective cohort with recorded injury outcomes, pre-register the modelling plan, retain trial-level data for reliability estimation, and validate externally before any clinical claim is advanced.

Conclusions

In a cross-sectional sample of 100 adult male basketball players, between-position differences in lower-extremity performance were limited: only BMI (η² = 0.130) and RSI (η² = 0.095) differed significantly, indicating modest positional specialisation at amateur-league level. A Random Forest classifier using field-based DeepSport performance metrics reproduced a composite lower-extremity screening classification with outstanding internal discrimination (AUC = 0.947; 0.930 ± 0.012 across seeds). This performance is substantially attributable to the fact that five of the thirteen predictors define the outcome: removing them reduces AUC to 0.683. The model should therefore be understood as an interpretable, automatable implementation of an existing screening rule, not as a validated injury-prediction tool. A conventional logistic regression performed comparably (AUC = 0.910, accuracy 87.0%), so the ensemble’s value lies chiefly in its ranking performance and its compatibility with per-athlete SHAP explanation. Limb asymmetry, BMI and 10-m sprint time were the dominant features across permutation, SHAP and Gini measures, supporting the routine assessment of bilateral asymmetry rather than reliance on jump-height testing alone; BMI, however, should be read as a proxy for absolute body size and joint loading rather than as an index of adiposity. With 4.6 events per variable and no external validation, these classification analyses are exploratory and internally validated only. Prospective studies with recorded injury outcomes and external validation are required before any predictive claim can be made.

Abbreviations

AUC

Area Under the Receiver Operating Characteristic Curve

BMI

Body Mass Index

CART

Classification and Regression Tree

CMJ

Countermovement Jump

COD

Change of Direction

CONSORT

Consolidated Standards of Reporting Trials

CV

Coefficient of Variation

EPV

Events Per Variable

ICC

Intraclass Correlation Coefficient

LSI

Limb Symmetry Index

ML

Machine Learning

NPV

Negative Predictive Value

PPV

Positive Predictive Value

RF

Random Forest

ROC

Receiver Operating Characteristic

RSI

Reactive Strength Index

SHAP

SHapley Additive exPlanations

SSC

Stretch-Shortening Cycle

STROBE

Strengthening the Reporting of Observational Studies in Epidemiology

TRIPOD

Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis

VIF

Variance Inflation Factor

Authors' contributions

EG: Writing – original draft, Writing – review and editing, Investigation, Methodology. SS: Formal Analysis, Investigation, Writing – original draft, Data curation, Methodology, Supervision, Visualization. GAT: Writing – original draft, Writing – review and editing, Investigation, Methodology, Supervision. GA: Visualization, Methodology, Writing – original draft, Writing – review and editing. MFB: Investigation, Methodology, Supervision, Writing – original draft.

Funding

The authors declare that they received no financial support for the research, authorship, and/or publication of this article.

Data availability

The datasets generated and/or analysed during the current study are not publicly available but are available from the corresponding author upon reasonable request. The dataset used for the analyses is de-identified and contains no direct or indirect identifiers. The complete analysis code, including data preparation, label construction, hyperparameter grid search, cross-validation, benchmarking, sensitivity analyses, permutation importance, TreeSHAP computation, and figure generation (random seed = 42), is also available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

This study was approved by the Non-Interventional Clinical Research Ethics Committee of Iğdır University (Decision No. 27/10, 06 May 2026; E-37077861-900-217860). Written informed consent was obtained from all participants. All participants were 18 years of age or older; no minors were enrolled and parental consent was therefore not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.