🤖 machine learning

Hybrid Machine Learning-Assisted Raman Spectroscopy with Generative Feature Augmentation for Pharmaceutical Identification

This paper introduces HyMLRaman, a hybrid framework that combines EfficientNet-B3 deep feature extraction with classical machine learning classifiers and DDPM-based generative augmentation to achieve high-accuracy identification of six pharmaceutical compounds from Raman spectra, even under limited-data conditions.

Original authors: Quach Thi Thai Binh, Ton Nu Quynh Trang, Thang B. Phan, Vu Thi Hanh Thu, Nguyen Tuan Hung

Published 2026-10-05

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Quach Thi Thai Binh, Ton Nu Quynh Trang, Thang B. Phan, Vu Thi Hanh Thu, Nguyen Tuan Hung

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the quiet corners of our food supply, water systems, and even clinical samples, traces of medicine often linger where they should not be. These pharmaceutical residues, remnants of drugs like antibiotics and pain relievers, pose a growing threat to public health and environmental safety. Detecting them quickly and accurately is a critical challenge for scientists and regulators. One powerful tool for this task is Raman spectroscopy, a technique that acts like a molecular fingerprint scanner. When light hits a molecule, it scatters in a unique pattern that reveals the specific chemical bonds holding that molecule together. This method is fast, non-destructive, and capable of identifying substances without touching them. However, the raw data from these scanners can be messy and variable, making it difficult for traditional computer programs to sort through the noise and find the true signal. To solve this, researchers are turning to a new approach that blends the pattern-recognition power of modern artificial intelligence with the reliable decision-making of older, proven statistical methods.

A team of researchers has developed a new system called HyMLRaman, designed to identify six common pharmaceutical compounds—amoxicillin, chloramphenicol, ciprofloxacin, tetracycline, ibuprofen, and paracetamol—directly from their Raman spectral data. Instead of relying on a single type of computer model, the team built a hybrid framework that first uses a deep neural network to translate the raw light patterns into a compact, high-level summary. They then feed this summary into a variety of classical machine-learning classifiers, such as support vector machines and random forests, to make the final identification. This two-step process proved highly effective. By converting the spectral data into a format that highlights the most important chemical features, the system achieved an accuracy of 96.31% in distinguishing between the six different drugs. This performance was significantly better than using the deep neural network alone, suggesting that combining the feature-extraction strength of modern AI with the decision-making stability of classical algorithms creates a more robust tool for real-world screening.

The researchers faced a common hurdle in scientific data collection: a limited amount of training material. To see if their system could still work when data was scarce, they introduced a generative component based on a diffusion model. This part of the system does not create new images of the drugs; instead, it learns the mathematical relationships between the existing data points and generates new, synthetic versions of the key features that describe the drugs. These synthetic features are then carefully filtered to ensure they are realistic and consistent with the original data before being added to the training set. The results showed that this augmentation was not a magic bullet that improved every scenario equally. It provided the most significant benefits when the training data was very small, particularly helping a specific type of classifier known as a nearest-neighbor algorithm to perform better. However, when the system had access to a full dataset, the synthetic additions offered little to no advantage, indicating that the technique is best used as a safety net for situations where data is sparse rather than a standard replacement for real measurements.

To prove that this framework could move beyond computer simulations and into practical use, the team built a working application called the Raman Pharmaceutical Analyzer. This software takes a raw Raman spectrum from an experimentally measured sample and instantly displays the predicted drug identity, the confidence level of that prediction, and a visual map showing which parts of the light pattern influenced the decision. When tested on independent samples of paracetamol, ciprofloxacin, ibuprofen, and tetracycline that the model had never seen before, the system correctly identified the compounds with confidence levels exceeding 98%. The visual maps confirmed that the system was focusing on the specific, chemically meaningful regions of the spectrum rather than guessing randomly. This demonstration suggests that the HyMLRaman framework offers a practical, interpretable, and reliable path forward for rapid screening of pharmaceutical residues, turning complex light patterns into clear, actionable information for protecting public health.

Technical Summary: HyMLRaman – Hybrid Machine Learning-Assisted Raman Spectroscopy with Generative Feature Augmentation

Problem Statement
The rapid and reliable identification of pharmaceutical residues (specifically amoxicillin, chloramphenicol, ciprofloxacin, tetracycline, ibuprofen, and paracetamol) is critical for public health and food safety. While Raman spectroscopy offers a non-destructive, molecule-specific analytical tool, practical application is hindered by spectral variability and the limitations of traditional chemometric approaches. Existing studies often rely on a single modeling paradigm—either classical machine learning (ML) requiring manual feature engineering or deep learning (DL) acting as a standalone classifier—failing to leverage the complementary strengths of both. Furthermore, many applications face challenges related to limited training data.

Methodology
The authors propose HyMLRaman, a hybrid framework that integrates deep spectral feature extraction with classical ML classifiers and generative feature augmentation. The methodology proceeds through the following stages:

Data Preprocessing and Image Construction:
- A dataset of 1,003 Raman spectral images was compiled from publicly available sources under comparable measurement conditions.
- Raw spectra underwent baseline correction, despiking, and smoothing.
- Spectra were reshaped into 128 × 128 grayscale images (converted to three-channel format) to serve as inputs for deep neural networks.
Deep Feature Extraction:
- Several state-of-the-art architectures were evaluated as feature extractors: CNNs (EfficientNet-B0, EfficientNet-B3, RegNetY-400M) and a Vision Transformer (ViT-B16).
- EfficientNet-B3 was identified as the optimal backbone, outperforming others in standalone classification (89.60% accuracy).
- The global average pooling layer of EfficientNet-B3 was used to generate 1536-dimensional embeddings, creating a unified feature space for downstream tasks.
Hybrid Classification:
- The 1536-D embeddings were fed into six classical ML classifiers: Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Logistic Regression (LR), Random Forest (RF), XGBoost, and a shallow Artificial Neural Network (ANN).
- Performance was evaluated using stratified 10-fold cross-validation.
- An embedding-level 1D CNN was also implemented as a neural baseline for comparison.
Generative Feature Augmentation (DDPM):
- To address limited-data scenarios, a Denoising Diffusion Probabilistic Model (DDPM) was trained within a PCA-reduced latent space (64 dimensions) of the EfficientNet-B3 embeddings.
- The DDPM generates class-conditional synthetic feature candidates.
- Filtering Strategy: Generated samples were not blindly added to the training set. Instead, they underwent a two-step filter:
  - Classifier Confidence: A logistic regression model verified that the generated sample's predicted class matched the conditioning class with probability $\ge 0.85$ .
  - Nearest-Neighbor Consistency: The sample had to lie within a distance threshold ( $d_{thr}$ ) of the original training distribution for that class.
- Only filtered samples were concatenated with original data for re-training classifiers in low-data ablation studies (25%, 50%, 75% data fractions).
Application Deployment:
- The trained EfficientNet-B3–SVM pipeline was integrated into a Raman Pharmaceutical Analyzer (PyQt interface) to demonstrate real-time inference on experimentally measured, independent test samples.

Key Results

Hybrid vs. Standalone Performance: The hybrid approach significantly outperformed standalone deep learning. While EfficientNet-B3 alone achieved 89.60% accuracy, the EfficientNet-B3–SVM configuration achieved 96.31% accuracy and a 96.36% macro-F1 score. Other hybrid models (ANN, LR, XGBoost) also surpassed 94% accuracy, while the embedding-level CNN baseline only reached 92.62%.
Classifier Comparison: SVM provided the most robust and balanced performance across all six pharmaceutical classes. KNN and RF showed slightly lower but competitive results.
Low-Data Ablation (DDPM): The impact of DDPM augmentation was selective and classifier-dependent, not universal:
- KNN showed the most consistent improvement in low-data settings (e.g., +0.795% accuracy at 50% data fraction), benefiting from the denser feature distribution.
- SVM showed marginal gains in specific low-data settings but generally remained stable.
- LR, XGBoost, and ANN often showed slight performance degradation or negligible changes when augmented, suggesting the synthetic features did not align well with their decision boundaries.
- At 100% data fraction, the augmentation provided minimal benefit, confirming its utility is primarily for sparse-data conditions.
Class-Level Discrimination: Confusion matrices and ROC-AUC analyses revealed that the hybrid models maintained high recall across all classes. SVM achieved class-wise AUC values between 0.995 and 0.999. The most significant residual confusion occurred between spectrally similar compounds like ibuprofen and paracetamol, particularly in KNN models.
Application Validation: The Raman Pharmaceutical Analyzer successfully identified independent experimental samples (e.g., paracetamol) with >98% confidence, providing interpretable spectral response maps.

Significance and Claims
The paper claims that HyMLRaman offers a practical and interpretable route for rapid pharmaceutical screening. Its primary contributions are:

Hybrid Architecture: Demonstrating that combining deep feature extractors (EfficientNet-B3) with classical margin-based classifiers (SVM) yields superior robustness and accuracy compared to pure deep learning or classical ML pipelines for Raman data.
Targeted Augmentation: Establishing that generative feature augmentation via DDPM in a latent space is a viable strategy for limited-data scenarios, but its efficacy is highly dependent on the downstream classifier (most beneficial for distance-based methods like KNN).
Feasibility: Proving that such models can be embedded into interactive workflows for real-time, on-site drug identification, bridging the gap between theoretical model performance and practical application.

The authors maintain a modest stance, noting that while the framework is promising, further validation on newly collected external samples is required before widespread practical deployment. The work does not claim to replace chemical evidence but rather to provide a rapid, high-confidence screening tool supported by interpretable spectral features.

Technical Summary: HyMLRaman – Hybrid Machine Learning-Assisted Raman Spectroscopy with Generative Feature Augmentation

More like this