Kolmogorov-Arnold Networks vs. Multi-Layer Perceptrons: A Rigorous Statistical Benchmark on Clinical Tabular Data

Rafał Wodziński; Paweł Szczerbiak

cs.LGNov 2, 2026

Kolmogorov-Arnold Networks vs. Multi-Layer Perceptrons: A Rigorous Statistical Benchmark on Clinical Tabular Data

Authors: Rafał Wodziński, Paweł Szczerbiak

Organizations: Sano Centre for Computational Medicine, Kraków, Poland · Department of Computer Science, Faculty of Computer Science and Mathematics, Cracow University of Technology, Kraków, Poland · Faculty of Computer Science, AGH University of Krakow, Kraków, Poland

Abstract

Kolmogorov-Arnold Networks (KANs) have recently been proposed as a promising alternative to Multi-Layer Perceptrons (MLPs). However, empirical evaluations on structured medical datasets remain scarce, and commonly reported metrics like AUROC can yield overly optimistic performance estimates on imbalanced clinical cohorts. In this study, we present a standardized benchmark comparing eleven architectures - nine KAN variants (FastKAN, WavKAN, ReLUKAN, ChebyKAN, TaylorKAN, GramKAN, JacobiKAN, LegendreKAN, and TabKAN), a tuned StandardMLP, and a Random Forest baseline - across seven heterogeneous medical tabular datasets spanning oncology, cardiology, nephrology, neurology, endocrinology, and obstetrics. Our experimental protocol employs a 3x5 Repeated Stratified K-Fold cross-validation scheme with nested Bayesian hyperparameter optimization and evaluates discrimination via Area Under the Precision-Recall Curve (AUPRC), calibration through the Brier Score, and computational efficiency. Formal equivalence testing is conducted using the Bayesian Correlated t-Test with a Region of Practical Equivalence (ROPE) at +/- 0.01. The results indicate that on datasets with severe class imbalance, dedicated KAN architectures (particularly TabKAN and ReLUKAN) can improve minority-class detection and achieve higher AUPRC than standard MLPs. Nevertheless, across pooled datasets, KAN variants demonstrate comparable overall predictive discrimination rather than a systematic advantage, while requiring 1.20x to 1.48x longer inference latency per sample and up to 12.5x more trainable parameters. Consequently, StandardMLP remains a strong baseline on the efficiency Pareto frontier, while KAN variants provide a viable alternative when precision on rare positive cases is the primary clinical objective. The code for reproducing the results is available at https://github.com/rafalwodzinski/kan_mlp_comparison.

Figures & tables

Dataset	$N$	$d$	Minority (%)	Task	Clinical Domain
Cervical Cancer [ 16 ]	858	32	6.4%	Binary	Gynaecology / Oncology
Cardiotocography [ 5 ]	2,126	21	8.3%	Multiclass	Obstetrics
Parkinson’s [ 32 ]	195	22	24.6%	Binary	Neurology
Pima Diabetes [ 41 ]	768	8	34.9%	Binary	Endocrinology
Breast Cancer [ 19 ]	569	30	37.3%	Binary	Oncology
Chronic Kidney [ 30 ]	400	24	37.5%	Binary	Nephrology

Table 1: Summary of Benchmark Medical Datasets

Figure 1: Distribution of per-fold Area Under the Precision-Recall Curve (AUPRC) scores for all eleven evaluated architectures across the seven medical datasets, shown as box-and-whisker plots (median, interquartile range, and 1.5-IQR whiskers). The results illustrate that while overall performance distributions overlap across balanced datasets, dedicated architectures such as TabKAN and FastKAN demonstrate competitive or superior retention of discriminative capacity on imbalanced cohorts (e.g., Cervical Cancer and Cardiotocography).

Panel of Figure 2 — Figure 2: Precision-Recall curves for representative clinical cohorts with high and moderate class imbalance, comparing neural architectures and Random Forest baseline across cross-validation folds.

Figure 3: Critical Difference (CD) Diagram for AUPRC constructed from average ranks across all seven benchmark datasets under the Nemenyi post-hoc test (

\alpha=0.05

). The horizontal axis represents the average rank of each model, where a lower rank indicates higher AUPRC. To interpret the diagram: the solid red line at the top visualizes the required critical difference distance (

CD=5.71

), and the thick gray line connects models whose pairwise rank differences are smaller than this threshold (i.e., they are not statistically significantly different). While TabKAN achieves the lowest average rank (

3.43

), the critical difference connects all evaluated models in a single clique. This indicates that rank differences across the heterogeneous benchmark do not cross the family-wise significance threshold, rendering all architectures statistically equivalent under this frequentist test.

Dataset	Best KAN	$P(\text{MLP}\succ)$	$P(\text{KAN}\succ)$	$P(\text{ROPE})$	$\bar{\Delta}$
Breast Cancer	TaylorKAN	<0.001	<0.001	1.000	+0.0014
Cardiotocography	WavKAN	<0.001	0.885	0.115	-0.0161
Chronic Kidney Disease	LegendreKAN	<0.001	<0.001	1.000	-0.0004
Heart Disease	GramKAN	0.081	0.116	0.803	-0.0008
Parkinson’s Disease	WavKAN	0.130	0.005	0.865	+0.0043
Pima Diabetes	TabKAN	0.289	0.343	0.368	-0.0016

Table 2: Bayesian Correlated

t

-Test Results (AUPRC). ROPE interval

=\pm 0.01

Figure 4: Distribution of Brier Scores across the repeated cross-validation folds. Lower scores indicate better probabilistic calibration, a critical requirement for safe clinical risk stratification.

Model	Grand Mean Brier Score
TabKAN	$0.0580\pm 0.0547$
StandardMLP	$0.0585\pm 0.0542$
WavKAN	$0.0591\pm 0.0569$
FastKAN	$0.0600\pm 0.0559$
ReLUKAN	$0.0606\pm 0.0558$
TaylorKAN	$0.0606\pm 0.0560$

Table 3: Grand Mean Brier Score across all datasets and folds (

\text{Grand Mean}\pm\text{Std}

, lower is better).

Figure 5: Pareto Frontier of grand mean AUPRC versus mean per-sample inference latency (log scale) across all eleven evaluated architectures. The empirical Pareto frontier is formed by StandardMLP (lowest latency:

0.0611

ms) and TabKAN (highest AUPRC:

0.9061

). Other KAN variants and baseline models remain within the dominated region, highlighting the operational trade-off between baseline execution speed and KAN minority-class retrieval performance.

Model	$\bar{t}_{\text{inf}}$ (ms)	Ratio vs. MLP	$\|\theta\|$ (Grand Mean)	$\bar{T}_{\text{train}}$ (s)
StandardMLP	0.0611	$1.00\times$	8,273	1.22
ReLUKAN	0.0735	$1.20\times$	73,906	2.04
FastKAN	0.0745	$1.22\times$	39,143	1.66
TabKAN	0.0752	$1.23\times$	56,990	2.04
TaylorKAN	0.0816	$1.34\times$	23,885	1.86
WavKAN	0.0826	$1.35\times$	103,347	2.23

Table 4: Computational Cost Summary (Grand Mean across all datasets and folds).

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Architecture	AUPRC	AUROC	MCC	F1-Score	Balanced Accuracy	Brier Score
TabKAN	$0.9061\pm 0.1262$	$0.9481\pm 0.0651$	$0.7890\pm 0.1959$	$0.8597\pm 0.1565$	$0.8858\pm 0.1062$	$0.0580\pm 0.0547$
StandardMLP	$0.9028\pm 0.1278$	$0.9544\pm 0.0611$	$0.7869\pm 0.1995$	$0.8543\pm 0.1549$	$0.8873\pm 0.1038$	$0.0585\pm 0.0542$
FastKAN	$0.9013\pm 0.1323$	$0.9480\pm 0.0630$	$0.7819\pm 0.1950$	$0.8624\pm 0.1393$	$0.8844\pm 0.1008$	$0.0600\pm 0.0559$
ReLUKAN	$0.8999\pm 0.1358$	$0.9464\pm 0.0642$	$0.7762\pm 0.2126$	$0.8581\pm 0.1470$	$0.8833\pm 0.1093$	$0.0606\pm 0.0558$
TaylorKAN	$0.8992\pm 0.1313$	$0.9470\pm 0.0634$	$0.7869\pm 0.2005$	$0.8637\pm 0.1340$	$0.8917\pm 0.1000$	$0.0606\pm 0.0560$
WavKAN	$0.8980\pm 0.1384$	$0.9471\pm 0.0646$	$0.7858\pm 0.1985$	$0.8597\pm 0.1520$	$0.8867\pm 0.1026$	$0.0591\pm 0.0569$

Appendix

Table 5: Summary of Predictive Performance (Grand Mean

\pm

Std across all datasets and folds)

Panel of Figure 6 — Figure 6: Gallery of cross-validation metric distributions for all evaluated architectures across seven medical datasets (from top to bottom): AUROC, MCC, macro-averaged F1-Score, and per-sample inference latency measured on Apple Silicon MPS hardware (log scale).

Panel of Figure 7 — Figure 7: Sample efficiency ablation curves illustrating AUPRC degradation as training data is reduced from $100\%$ to $10\%$ across six clinical cohorts (left to right, top to bottom): Cervical Cancer, Cardiotocography, Parkinson’s Disease, Pima Diabetes, Breast Cancer, and Heart Disease. Shaded bands denote $\pm 1$ standard deviation across folds.

Panel of Figure 8 — Figure 8: Training and validation loss trajectories across four distinct clinical datasets, comparing baseline StandardMLP against representative KAN variants (separate columns). Rows from top to bottom: Cervical Cancer ( $6.4\%$ positive prevalence), Cardiotocography ( $8.3\%$ minority class), Heart Disease, and Breast Cancer.

Panel of Figure 9 — Figure 9: Confusion Matrix Gallery comparing StandardMLP against representative KAN variants (separate columns). Rows from top to bottom: Cervical Cancer ( $6.4\%$ positive prevalence), Cardiotocography ( $8.3\%$ minority class), Pima Diabetes, and Breast Cancer.

Kolmogorov-Arnold Networks

Kolmogorov-Arnold Networks (KANs) learn compositional functions through trainable one-dimensional edge transformations for approximation and interpretable modeling.

See all 126 papers

Is the Statistical Advantage Worth the Cost? An Empirical Comparison of KANs and MLPs for Structured Data Classification
Jul 15, 2026Matthew Steven P. Toledo, Justine Raphael H. Jacinto, Vivekjeet Singh Chambal +3Multilayer Perceptrons Tabular Classification
Institute of Computer Science, University of the Philippines Los Baños, Philippines · Machine Learning and Artificial Intelligence Applications Lab, University of the Philippines Los Baños, Philippines
KANLib -- A Modular, Extensible and Fast Kolmogorov-Arnold Network Implementation
Jun 16, 2026Julian Hoever, Gregor SchieleKolmogorov-Arnold Networks Modular Neural Networks
Intelligent Embedded Systems University of Duisburg-Essen Duisburg, Germany
Kolmogorov-Arnold Networks for Personal Context Recognition on ExtraSensory
Oct 4, 2026Hoang-Thang TaKolmogorov-Arnold Networks
University of Information Technology, Ho Chi Minh City, Vietnam · Vietnam National University Ho Chi Minh City, Ho Chi Minh City, Vietnam

Jul 15, 2026J/K move · Enter open · S save

Is the Statistical Advantage Worth the Cost? An Empirical Comparison of KANs and MLPs for Structured Data Classification

Matthew Steven P. Toledo, Justine Raphael H. Jacinto, Vivekjeet Singh Chambal, Rodolfo C. Camaclang+2

Institute of Computer Science, University of the Philippines Los Baños, Philippines · Machine Learning and Artificial Intelligence Applications Lab, University of the Philippines Los Baños, Philippines

Multilayer Perceptrons Tabular Classification Kolmogorov-Arnold Networks Tabular ML

Open paper AskPDF arXiv

This study presents an empirical benchmarking comparison between Kolmogorov-Arnold Networks (KANs) and Multi-Layer Perceptrons (MLPs) on structured tabular classification tasks. Motivated by the growing interest in KANs as an alternative function-approximating architecture, we evaluate their out-of-the-box performance on twelve publicly available datasets spanning binary, multiclass, multilabel, and ordinal problems. Both models were trained under standardized preprocessing, architecture, and fixed hyperparameter settings, with performance assessed using test accuracy and F1-Score, paired hypothesis testing, and effect size analysis. Results show that KANs statistically outperform MLPs in binary and multiclass domains and achieve a significant aggregate advantage across all datasets. However, the observed medium effect size (d = -0.46) raises an important cost-benefit consideration: while KANs offer superior generalization through adaptive spline-based mappings, this advantage comes with substantially higher parameter and computational complexity relative to the MLP baseline. These findings suggest KANs are the preferred choice for high-precision applications, while MLPs remain a robust and efficient option for resource-constrained environments. Future work should extend this analysis to additional data modalities to further refine these architectural selection criteria.

Abstract

Kolmogorov-Arnold Networks

Explore similar work

Is the Statistical Advantage Worth the Cost? An Empirical Comparison of KANs and MLPs for Structured Data Classification

KANLib -- A Modular, Extensible and Fast Kolmogorov-Arnold Network Implementation

Kolmogorov-Arnold Networks for Personal Context Recognition on ExtraSensory