Strategies to Enrich Electrochemical Sensing Data with Analytical Relevance for Machine Learning Applications: A Focused Review

Mijeong Kang, Donghyeon Kim, Jihee Kim, Nakyung Kim, Seunghun Lee2024 · DOI

TL;DR

This review explores strategies to enhance electrochemical sensing data for machine learning (ML) applications. By enriching data with analytical relevance, researchers can improve the accuracy and efficiency of ML models in electroanalytical science, particularly for complex real-world samples.

What Problem Does This Paper Address?

Electrochemical sensors offer a powerful tool for detecting and quantifying substances, but their application in complex matrices (like biological fluids or environmental samples) is often hindered by challenges in distinguishing target analytes from interfering substances. While machine learning (ML) holds immense promise for analyzing electrochemical data, its effectiveness is fundamentally limited by the quality and richness of the input data. Traditional electrochemical data, while informative, may lack the necessary analytical context to enable ML models to accurately identify and quantify targets, especially at low concentrations or in the presence of numerous non-target compounds. This paper addresses the critical need to develop strategies that generate more analytically relevant electrochemical data, thereby maximizing the potential of ML in electroanalytical science.

Methods

The paper reviews five key strategies for enriching electrochemical sensing data for ML applications:

  • Strategy I: Multi-Electrode Systems with Differential Interactions: This involves using an array of electrodes made from different materials or modified differently. Each electrode interacts uniquely with target substances, generating a diverse dataset that forms an "electrochemical fingerprint" for better identification.
  • Strategy II: Leveraging Multiple Electrochemical Techniques: Instead of relying on a single measurement technique (e.g., cyclic voltammetry), this strategy combines data from multiple techniques (e.g., CV, differential pulsed voltammetry, chronoamperometry). This multi-dimensional data provides a more comprehensive view of the analyte's behavior, improving ML model performance.
  • Strategy III: Dissecting Signals for Trace-Level Detection: This focuses on extracting subtle information from electrochemical signals, particularly for detecting analytes at very low concentrations. Techniques like anodic stripping voltammetry (ASV) are employed, and sophisticated feature extraction from the signal shape is used to capture trace-level information. Data augmentation techniques, such as random perturbation, are also discussed to increase dataset size and mitigate overfitting.
  • Strategy IV: Designing Systems Perturbed Differently by Targets: This approach involves creating electrochemical systems where the redox reaction is influenced differently by various target substances. By observing these unique "fingerprints" in the electrochemical response, ML can be used to discriminate between analytes, even when they share similar chemical properties or when using a single sensing electrode.
  • Strategy V: Using Raw Voltammetric Data for Reduced Processing: This strategy aims to simplify the data processing pipeline, making electrochemical sensors more suitable for point-of-care (POC) applications with limited computational resources. By feeding raw, unprocessed data directly into ML models, the processing cost is reduced, enabling faster analysis on portable devices.

Datasets

The paper discusses several studies that utilize various datasets for training and evaluating ML models. These datasets are typically generated from electrochemical measurements of specific analytes in different matrices.

  • Antibiotic Detection in Milk: A dataset was created using cyclic voltammograms from milk samples with and without various antibiotics at different concentrations. This dataset was used to train ML models for antibiotic identification.
  • Chloramphenicol (CAP) and Metronidazole (MNZ) Detection: Data from CV, DPV, and CA measurements of CAP and MNZ mixtures at varying concentrations were used. Features extracted from these signals were used to predict CAP concentration.
  • Copper Ion (Cu²⁺) Detection in Cell Culture Media: Anodic stripping voltammetry (ASV) curves were collected for cell culture media with different Cu²⁺ concentrations. These curves, along with extracted features, were used to train ML models for Cu²⁺ quantification.
  • Volatile Organic Compound (VOC) Discrimination: Cyclic voltammograms were recorded for various VOCs in a designed electrochemical system. These voltammograms were used to train ML models for VOC identification.
  • Dopamine (DA), Ascorbic Acid (AA), and Uric Acid (UA) Detection: Raw square-wave voltammetry (SWV) data from mixtures of these analytes were used to train ML models for detecting DA in the presence of interferents.

The paper emphasizes the importance of dataset size and class balance for accurate ML predictions.

Key Insights and Ideas

  • Data is King (and Queen): The paper strongly emphasizes that advanced ML algorithms are only as good as the data they are trained on. Simply applying ML to noisy or insufficient electrochemical data will yield poor results.
  • Beyond the Single Signal: The future of electroanalytical ML lies in integrating information from multiple sources – different electrodes, different measurement techniques, and subtle signal features – to create a richer, more informative dataset.
  • "Fingerprinting" Analytes: The concept of creating unique "electrochemical fingerprints" by combining diverse data sources is a powerful strategy for complex identification tasks.
  • Smart Design for Smart Analysis: Designing electrochemical systems that inherently produce differential responses to analytes (Strategy IV) can simplify the ML task and reduce the need for complex post-processing.
  • POC Needs Simplicity: For point-of-care applications, reducing computational load is crucial. Using raw data and efficient ML models (Strategy V) is a promising direction.
  • Data Augmentation is Crucial: For limited datasets, techniques like random perturbation can significantly boost performance and mitigate overfitting, especially for trace analysis.

Key Results

The paper highlights several successful applications of these strategies:

  • Antibiotic Identification: A multi-electrode system achieved classification accuracies ranging from 0.8 to 1.0 for 5 antibiotics and 0.55 to 1.0 for 15 antibiotics, demonstrating the effectiveness of diverse electrode responses.
  • CAP Detection: Combining CV, DPV, and CA data improved prediction accuracy for CAP concentration to 0.95, significantly outperforming individual techniques (0.80 for CV, 0.69 for DPV, 0.36 for CA).
  • Cu²⁺ Quantification: Support vector machines (SVM) were found to be the most accurate model for predicting Cu²⁺ concentrations using features extracted from ASV curves.
  • VOC Discrimination: Linear discriminant analysis (LDA) effectively discriminated between different VOCs based on their unique electrochemical fingerprints.
  • POC Biosensing: A neural network model using raw SWV data achieved classification accuracies over 95% for dopamine detection, with significantly reduced memory usage and energy consumption compared to processed data.

Limitations and Future Work

  • Data Quality and Quantity: The paper repeatedly stresses that the performance of ML models is highly dependent on the quality and quantity of the experimental data. Generating large, well-characterized datasets remains a challenge.
  • Overfitting: With high-dimensional data, there's an increased risk of overfitting, where models perform well on training data but poorly on new, unseen data. Strategies like data augmentation and careful feature selection are crucial to mitigate this.
  • Generalizability: While specific strategies show promise, ensuring the generalizability of ML models across different sensor platforms, matrices, and analyte combinations is an ongoing area of research.
  • Computational Resources: For POC applications, further work is needed to optimize ML models for even lower power consumption and memory footprints without sacrificing accuracy.
  • Standardization: The development of standardized protocols for data acquisition, feature extraction, and model validation would accelerate the adoption of ML in electroanalytical science.

The authors envision a future where ML is seamlessly integrated into electrochemical analysis, enabling more sophisticated, autonomous, and efficient sensing systems.

Summary updated

BibTeX
@misc{kang2024strategies,
  author = {Mijeong Kang and Donghyeon Kim and Jihee Kim and Nakyung Kim and Seunghun Lee},
  title = {{Strategies to Enrich Electrochemical Sensing Data with Analytical Relevance for Machine Learning Applications: A Focused Review}},
  year = {2024},
  doi = {10.3390/s24123855},
}

Loading related content...

Figure