Work

A Comparative Study of Neural Architecture Search Methods for Medical Data Classification

Year: 2026
Type: article
Abstract: Medical datasets can be small or imbalanced, making the selection of suitable neural network architectures challenging. NeuroEvolution of Augmenting Topologies (NEAT) can automate the evolution of network architectures. We tested six NEAT variants: NEAT, FS-NEAT, NEAT-DBM, HyperNEAT, C-NEAT and L-NEAT. This study aims to assess the relative benefits of these variants under a common experimental protocol. We conducted 15 runs of a repeated stratified hold-out on 20 publicly available diverse medical datasets. All algorithms had a common configuration, common implementation, nominal budget and no hyperparameter tuning. Tuning was omitted; hence, no algorithm benefited from a greater tuning effort. We chose test macro F1 as the main metric, and the Friedman test and Holm procedure for statistical analysis. Macro F1 was selected to give equal weight to all classes across datasets with different class distributions. L-NEAT, NEAT-DBM and NEAT had the best average ranks; however, they did not differ significantly among themselves. In contrast, HyperNEAT was significantly worse than NEAT, NEAT-DBM, L-NEAT and C-NEAT. FS-NEAT was significantly worse than L-NEAT, NEAT-DBM and NEAT. C-NEAT was significantly worse than L-NEAT. In the sensitivity analysis, five of the eight significant differences remained stable, whereas the NEAT–FS-NEAT, C-NEAT–L-NEAT and C-NEAT–HyperNEAT differences were sensitive to the removal of one dataset from each pair with overlapping data. L-NEAT was considerably slower than NEAT and NEAT-DBM without a significant difference in macro F1. When computational resources are limited, NEAT and NEAT-DBM might be a reasonable alternative to L-NEAT.
Language: English
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W7236584656
FWCI:
Cites: 37
Cited by:
Related to: 10