# Predictive modeling of flavonoid efficacy against esophageal carcinoma: a comprehensive approach

## TL;DR

This study uses machine learning to predict effective doses of natural flavonoids for treating esophageal cancer, offering a new approach to cancer therapy.

## Contribution

A novel predictive framework combining machine learning and flavonoid research to optimize cancer treatment strategies.

## Key findings

- A decision tree model achieved 87.38% accuracy in predicting optimal flavonoid doses for cancer cell viability.
- Seven flavonoids were evaluated for their efficacy against esophageal carcinoma cells.
- The model enables reliable predictions across diverse cancer cell lines.

## Abstract

Esophageal carcinoma poses a significant health challenge, particularly due to its notably high prevalence in East Asia, which underscores the urgent need for innovative treatment strategies. Natural flavonoids are polyphenolic compounds with significant potential to interact with cancer cell pathways; however, their therapeutic application remains constrained by the lack of comprehensive data and consistent experimental evidence. The current study aims to present a comprehensive framework for evaluating the anticancer potential of natural flavonoids, including seven flavones—luteolin, apigenin, chrysin, acacetin, cirsiliol, baicalein and eupatilin—and six flavonols—kaempferol, quercetin, galangin, myricetin, casticin, and gossyptin—against human esophageal carcinoma cells. Extensive experimental data from twenty-two research studies focusing on natural flavonoids and esophageal carcinoma cells were extracted based on specific inclusion criteria that emphasized dose, time, and cell viability. Seven machine learning models—K-Nearest Neighbors, Support Vector Machine, Logistic Regression, Decision Tree (DT), Random Forest, AdaBoost and XGBoost—were employed to predict the optimal dose and time required for flavonoids to achieve 50% cancer cell viability. Hyperparameters were fine-tuned for each algorithm, followed by 5-fold cross-validation on the train dataset to identify the most accurate predictive model. The performance of each algorithm was validated through rigorous hyperparameter optimization and validation on an independent test dataset. The DT model was identified as the most accurate predictor, leading to the development of a simplified model with low complexity that achieved an accuracy of 87.38%. This reliable model enables prediction of effective dosing parameters across diverse esophageal carcinoma cell lines. In conclusion, present study offers a robust yet accessible predictive tool for optimizing treatment strategies against esophageal carcinoma. By combining machine learning with natural compound research, this work exemplifies a transformative approach in oncology, accelerating the development of effective cancer therapies.

The online version contains supplementary material available at 10.1038/s41598-025-23689-2.