--- license: cc-by-nc-sa-4.0 language: - en library_name: keras tags: - regression - tensorflow - keras - colourimetric-sensing - biosensing - cholesterol - research-demo --- # Cholesterol Concentration Predictor — ANN Model ## Model summary This model repository accompanies research on the use of artificial neural networks (ANNs) and an Au@CeO₂ (gold–cerium oxide) nanozyme-based colourimetric sensor for cholesterol estimation. The corresponding **Gradio application** enables users to compare ANN predictions with a linear-regression baseline using mean RGB colour measurements. **Publication:** Govindharaj, P., Alungal, N., Emerald, C. S. and Kannan, S. (2026). *Au@CeO₂ nanozyme based smart colourimetric sensor for cholesterol: A neural network powered point of care solution model.* **Biochemical Engineering Journal**, **225**, 109908. DOI: [10.1016/j.bej.2025.109908](https://doi.org/10.1016/j.bej.2025.109908). **Links:** [Associated publication](https://doi.org/10.1016/j.bej.2025.109908) · [Interactive Gradio Space](https://huggingface.co/spaces/CaxtonEmeraldS/CholesterolConcentrationPredictor) · [Researcher's website](https://caxtonemerald-s.github.io) ## Intended uses - Research and educational exploration of ANN-based colourimetric regression. - Comparing predictions under different neural-network configurations **in the companion application**, to the extent that their trained files are available in its archive. - Reproducing and examining the method when the original measurement protocol, preprocessing and training artefacts are available. **Not intended for diagnosis, screening, monitoring, treatment selection or any other clinical use.** This repository does not establish medical-device validation. ## Data and features The application uses **three numerical inputs**: mean red, green and blue intensities. Each is expected to be in the range **0–255**. These values should come from appropriately collected colourimetric measurements, rather than arbitrary photographs or uncontrolled lighting conditions. The application code uses this transform for inference: ```python x = [[R / 256.0, G / 256.0, B / 256.0]] ``` ## Architecture and model coverage The application describes a **3–N–1** feedforward ANN: three RGB features, one hidden layer of variable width `N`, and a single output. Approximately **2,500 trained model configurations**, varying activation function, random seed and hidden-neuron count. The Gradio Space locates its model files under: ```text Models.zip └── Models/ └── / └── / └── model_.keras ``` Configuration from the repository name or the publication's best-performing model can be inferred. ## Baseline and evaluation The application implements a linear-regression comparator using normalised RGB inputs: ```python linear_raw = ( 1.9221 * (R / 256.0) - 1.3817 * (G / 256.0) + 1.4058 * (B / 256.0) - 0.1318 ) linear_mM = max(0.0, linear_raw) * 50.0 ``` The publication's abstract reports: | Measure | Study-reported value | |---|---:| | Best validation MSE | 0.0577 (hard shrink, 13 hidden neurons) | | Lowest training MSE | 0.0637 (SELU, 7 hidden neurons) | | Median training MSE | 0.0896 | | Median validation MSE | 0.0769 | | Reported colourimetric linearity range | 0–50 mM | ## How to use **Recommended:** open the [companion application](https://huggingface.co/spaces/CaxtonEmeraldS/CholesterolConcentrationPredictor), supply calibrated mean RGB values, choose an available activation, seed and hidden-neuron count, and compare the outputs. **For programmatic use:** examine the repository's actual files and loading instructions for the specific saved Keras model. The general TensorFlow/Keras pattern is: ```python import numpy as np import tensorflow as tf model = tf.keras.models.load_model("path/to/model.keras", compile=False) x = np.array([[R / 256.0, G / 256.0, B / 256.0]]) raw = float(np.asarray(model.predict(x, verbose=0)).reshape(-1)[0]) predicted_mM = max(0.0, raw) * 50.0 ``` Replace `path/to/model.keras` with the downloaded model path. ## Limitations, safety and licence - Experimental results depend on sample preparation, sensor chemistry, calibration, lighting, camera characteristics, image processing and domain shift. - Predictions are **not** calibrated confidence intervals and should not be used as medical measurements. - Clamping negative outputs to zero does not correct model error or guarantee validity. - The publication reports a range of 0–50 mM; extrapolation and comparisons outside its study conditions require separate validation. ## Citation Govindharaj, P., Alungal, N., Emerald, C. S. and Kannan, S. (2026). Au@CeO₂ nanozyme based smart colourimetric sensor for cholesterol: A neural network powered point of care solution model. *Biochemical Engineering Journal*, 225, 109908. https://doi.org/10.1016/j.bej.2025.109908 **Contact:** [Caxton Emerald S — professional website](https://caxtonemerald-s.github.io) · [ORCID](https://orcid.org/0000-0002-7763-3987)