Ethnobotanical risk assessment of visually similar medicinal plants: a machine learning approach to identifying species prone to misidentification in traditional medicine

Abstract

Background: Plant misidentification in traditional medicine is an underreported public health risk because visually similar medicinal species are often confused, leading to therapeutic failure or toxicity. Previous work addressed medicinal plant image classification or documented misidentification separately, but none linked machine-learning confusion patterns with pharmacological risk profiling.

Methods: This study trained a Convolutional Neural Network (CNN) and a Siamese Network with triplet loss on 20,000 leaf images from 20 medicinal plant species, sourced from two Bangladeshi repositories. The Siamese Network extracted 128-dimensional embeddings to compute Euclidean distances for all 190 species pairs. Misclassification rates were used to assign risk categories, and Pearson correlation tested the link between embedding proximity and error rates.

Results: The CNN achieved 88.45% overall accuracy. The Siamese Network identified 14 critical pairs (≥6% misclassification) and 7 high-risk pairs (4-5%). The most confusable pair was Justicia adhatoda misidentified as Terminalia arjuna (14%), followed by J. adhatoda as Centella asiatica (12%), Kalanchoe pinnata as T. arjuna (11%), and K. pinnata as Mikania micrantha (10%). A significant negative correlation (r = −0.304, p < 0.0001) confirmed that closer embeddings correspond to higher error rates.

Conclusions: The 21 high-risk pairs involve pharmacologically incompatible compounds: vasicine from J. adhatoda acts as a uterotonic and abortifacient; bufadienolides from K. pinnata cause cardiac glycoside‑like toxicity; cardenolides from Calotropis gigantea are cardiotoxic; and Hibiscus rosa‑sinensis has antifertility effects. These findings support prioritizing field training, visual guides, and digital tools to reduce plant misidentification in herbal medicine.

Keywords: Medicinal plant misidentification; Ethnobotany; Deep learning; Siamese Network; Traditional medicine safety; Pharmacological risk; Plant morphology; Visual Similar