Machine Learning for Heavy Metal Pollution Status Classification Using Spectroscopic and Electrochemical Data: A Systematic Review

2.1. Research design

This study employed a systematic literature review design to identify, evaluate, and synthesize peer-reviewed research articles that apply machine learning (ML) or chemometric algorithms for heavy metal pollution status classification based on spectroscopic and/or electrochemical data in environmental matrices. The review was structured around the P-E-O (Population, Exposure, Outcome) framework (Methley et al., 2014; Schardt et al., 2007), as illustrated in Figure 2.

Figure 2. P-E-O conceptual framework.

Figure 2 shows the schematic of the P-E-O framework applied in this review.

2.2. Eligibility criteria

Studies were included or excluded based on the criteria summarized in Table 1.

2.3. Information sources

Searches were conducted in the following electronic databases:

These databases were selected due to their comprehensive coverage of peer-reviewed literature in environmental science, analytical chemistry, materials science, and engineering disciplines relevant to heavy metal detection and machine learning applications. No additional databases (e.g., Web of Science, PubMed, IEEE Xplore) were used. The final search was performed on 6 May 2026. No filters for publication status were applied beyond the date range (2015–2026). Reference lists of included studies and relevant review articles were manually screened to identify additional eligible studies (snowballing). Table 2 presents the number of records initially identified from each database.

2.4. Search strategy

A systematic search was conducted on 6 May 2026 using three electronic databases: Scopus, SpringerLink, and ScienceDirect. The same core search string was applied across all databases, with minor syntactic adjustments to accommodate each platform. The search combined keywords related to three main concepts: (1) heavy metals, (2) spectroscopic/electrochemical techniques, and (3) machine learning/chemometrics, using Boolean operators (AND, OR). The base search string was:

Additional filters were applied uniformly: peer-reviewed research articles, English language, and publication year >2014 (or 2015–2026 for ScienceDirect). The search was performed by two independent reviewers; discrepancies were resolved by consensus. Results were exported to reference management software for duplicate removal. The number of records retrieved per database is reported in Table 2 (Section 2.3).

2.5. Study selection process

Study selection followed a four-stage process, as illustrated in Figure 1 (Section 3.1). The process was conducted independently by two reviewers, with disagreements resolved through consensus or consultation with a third reviewer.

Stage 1: Identification. A total of 825 records were identified from the three databases (Scopus: 133; SpringerLink: 681; ScienceDirect: 11) on 6 May 2026. All records were exported to reference management software, where duplicates were removed. After duplicate removal, the remaining records proceeded to the screening stage.

Stage 2: Screening (title and abstract). The titles and abstracts of all unique records were screened against the eligibility criteria ( Table 1). Records that clearly did not meet the inclusion criteria were excluded. When eligibility could not be determined from the title and abstract alone, the full text was retained for the next stage. This stage yielded 154 records considered potentially relevant and thus included for full-text review.

Stage 3: Eligibility (full-text screening). Full texts of the 154 potentially eligible records were retrieved and assessed independently by two reviewers against the full eligibility criteria. A total of 143 records were excluded at this stage, primarily because the study area was not relevant to the review’s focus (e.g., target analyte not a heavy metal, matrix not environmental, outcome not classification of pollution status, or no machine learning algorithm applied). Reasons for exclusion were documented for each excluded full text.

Stage 4: Included. After full-text screening, only 11 studies met all inclusion criteria and were included in the final qualitative synthesis.

The PRISMA flow diagram ( Figure 1, Section 3.1) provides a complete visual summary of the selection process, including the number of records at each stage and reasons for exclusion.

2.6. Data extraction

Data were extracted from each included study using a standardized data extraction form developed in Microsoft Excel. The form was piloted on three randomly selected studies and refined accordingly. Data extraction was performed by one reviewer and independently verified by a second reviewer. Discrepancies were resolved through discussion.

The extracted information included bibliographic details, study characteristics, target heavy metal(s), matrix type, concentration range, regulatory threshold, spectroscopic or electrochemical technique, preprocessing methods, machine learning algorithms, validation approach, classification performance metrics (accuracy, precision, recall, F1-score, AUC), and any integration of multiple data types.

Due to space limitations, the full data extraction results are not presented within this article. However, the complete dataset, including all extracted variables and intermediate calculations, is available in the supplementary repository (see Data Availability Statement). A summary of the extracted data is presented in Tables 4–12 (Section 3).

2.7. Data items

Data were extracted from each included study covering the following items:

No assumptions were made for missing or unclear data; items not reported were recorded as “not reported”. A summary of the extracted data is provided in Tables 3–12 (Section 3).

2.8. Study risk of bias assessment

Risk of bias was assessed independently by two reviewers using a tailored tool adapted from the Joanna Briggs Institute (JBI) checklist and QUADAS-2 (Moola et al., 2020; Whiting et al., 2011). Six domains were evaluated:

Each domain was rated as “low risk”, “high risk”, or “unclear risk”. Studies were not excluded based on bias scores; the assessment was used to inform the narrative synthesis and sensitivity analysis. Results are summarized in Table 12 and Figure 7 (Section 3.8).

2.9. Effect measures

Due to anticipated heterogeneity in study designs, target analytes, matrices, and classification tasks, a meta-analysis was not performed. Effect measures were summarized narratively. For classification performance, the following metrics were extracted when reported:

When multiple models were reported, the best-performing model based on accuracy or AUC was prioritized. For studies that compared integrated vs. single-modality approaches, the improvement in accuracy (percentage points) was calculated. A summary of performance metrics is presented in Table 10 and Figures 5–6 (Section 3.6).

2.10. Synthesis methods

A narrative synthesis approach was adopted (Popay et al., 2006). The synthesis was structured around the four research questions and organized thematically:

Subgroup analyses were planned by matrix type (water vs. soil), data type (spectroscopy vs. electrochemistry), and algorithm family (traditional ML vs. deep learning vs. ensemble). A sensitivity analysis was conducted by excluding studies with high risk of bias to assess the robustness of the findings. Publication bias was not formally assessed because of the narrative nature of the synthesis and heterogeneity of outcome measures.

3.1. Study selection

The systematic search yielded a total of 825 records from three databases: Scopus (n = 133), SpringerLink (n = 681), and ScienceDirect (n = 11). After duplicate removal, 567 unique records remained. Title and abstract screening excluded 413 records, leaving 154 records for full-text assessment. Full-text screening resulted in the exclusion of 143 records, primarily because the study area was not relevant to the review’s focus (e.g., target analyte not a heavy metal, matrix not environmental, outcome not pollution status classification, or no machine learning algorithm applied). A total of 11 studies met all eligibility criteria and were included in the final qualitative synthesis. The PRISMA flow diagram is presented in Figure 1.

3.2. Characteristics of included studies

A summary of the 11 included studies is presented in Table 3. The studies were published between 2019 and 2025, with the majority appearing after 2022. The corresponding authors were primarily from China (n = 5), the United States (n = 3), South Korea (n = 2), and India (n = 1). All studies were peer-reviewed original research articles published in English.

3.3. Distribution by year and geographic origin

The number of included studies increased over time, as shown in Figure 2. Only one study was published before 2020 (Dean et al., 2019), followed by two studies in 2022 (Hajzus et al., 2022; Park et al., 2022), three in 2023 (Chen et al., 2025; Maity et al., 2023; Wei et al., 2023), two in 2024 (Hu et al., 2024; Kailasam et al., 2024), and three in 2025 (Lahari et al., 2025; Qi et al., 2025; Valle et al., 2025). The annual growth rate reflects increasing research interest in applying machine learning for heavy metal pollution classification.

Geographically, most studies originated from Asia (China, South Korea, India) and North America (United States). The geographic distribution is presented in Figure 3.

Figure 3. Geographic distribution of included studies by corresponding author affiliation.

3.4. Population characteristics: heavy metals and sample matrices

The most frequently targeted heavy metals were lead (Pb, 54.5%, n = 6), cadmium (Cd, 45.5%, n = 5), and mercury (Hg, 45.5%, n = 5), followed by copper (Cu, 36.4%, n = 4), arsenic (As, 27.3%, n = 3), chromium (Cr, 18.2%, n = 2), and nickel (Ni, 18.2%, n = 2). The frequency distribution is presented in Table 4.

Regarding sample matrices, water dominated the included studies (72.7%, n = 8), including surface water, tap water, lake water, seawater, and wastewater. Soil was the second most common matrix (18.2%, n = 2). No studies on sediment matrices met the inclusion criteria. The distribution of sample matrices is presented in Table 5.

3.5. Exposure characteristics: data types and algorithms

Among the 11 included studies, 8 (72.7%) used electrochemical techniques, while 4 (36.4%) used spectroscopic techniques (note: one study, Chen et al., used fluorescence spectroscopy). No study integrated both data types. The distribution of data types is presented in Figure 4.

Figure 4. Distribution of data types across included studies.

Spectroscopic techniques used included vis-NIR spectroscopy (Hu et al., 2024; Qi et al., 2025), SERS (Park et al., 2022; Wei et al., 2023), and 3D fluorescence spectroscopy (Chen et al., 2025). Details are provided in Table 6.

Electrochemical techniques included cyclic square wave voltammetry (CSWV), differential pulse voltammetry (DPV), electrochemical impedance spectroscopy (EIS), and graphene field-effect transistor (GFET). Details are provided in Table 7.

Machine learning algorithms: The most frequently used algorithms were Support Vector Machine (SVM, 45.5%, n = 5), Random Forest (RF, 36.4%, n = 4), and Artificial Neural Network (ANN/MLP, 36.4%, n = 4). Deep learning algorithms (CNN, LSTM) appeared in 5 studies and demonstrated the highest performance (AUC > 0.99). The frequency distribution is presented in Table 8.

Preprocessing and feature extraction methods commonly included baseline correction, normalization (Z-score, Min-Max, area), PCA for dimensionality reduction, and Savitzky-Golay smoothing. A summary is provided in Table 9.

3.6. Outcome characteristics: classification performance

All 11 studies reported classification outcomes. Seven studies performed binary classification (contaminated vs. non-contaminated or above vs. below threshold), while four studies performed multi-class classification (metal type identification or concentration level classification). The performance metrics are summarized in Table 10.

The best-performing models were deep learning algorithms (CNN, LSTM, ALSTM-FCN), achieving AUC values up to 0.999 and accuracy up to 100%. Ensemble methods (BalancedRandomForest) also demonstrated robust performance for imbalanced soil datasets. A comparison of accuracy across studies is presented in Figure 5, and representative ROC curves are shown in Figure 6.

Figure 5. Best classification accuracy reported by each included study.

Figure 6. Representative micro-average ROC curves for the 11-class seawater dataset (11-SW).

The ALSTM-FCN model achieved an AUC of 0.999.

3.7. Integration of spectroscopic and electrochemical data

None of the 11 included studies (0%) integrated spectroscopic and electrochemical data. All studies used only a single data modality: spectroscopy alone (n = 4, 36.4%) or electrochemistry alone (n = 7, 63.6%). Therefore, no evidence was available to assess whether data integration improves classification accuracy compared to single-modality approaches. The findings are summarized in Table 11.

3.8. Risk of bias assessment

The risk of bias assessment for the 11 included studies is summarized in Table 12 and visualized in Figure 7. Most studies showed low risk of bias across the six domains. Sample representativeness, reference standard, ML model validation, and reporting of performance metrics were generally adequate. Issues were identified in some studies regarding the justification of classification thresholds (e.g., Valle et al. (2025) did not specify a threshold) and handling of confounding factors (e.g., pH, temperature, matrix effects were sometimes not addressed). No study was excluded based on risk of bias.

Figure 7. Risk of bias summary for 11 included studies.

3.9. Summary key findings

From 825 initial records ( Table 2), only 11 studies ( Table 3) met all inclusion criteria. Key findings include:

4.1. Limitations of the evidence included in the review

The body of evidence has several intrinsic limitations that affect the strength and generalisability of our conclusions.

First, heterogeneity in classification tasks and thresholds. Some studies defined binary contamination based on WHO drinking water guidelines (e.g., Pb > 10 μg/L), while others used national or regional standards (e.g., Chinese soil quality standards, EU guidelines). Multi-class classification tasks varied widely: some distinguished metal types (Pb vs. Cd vs. Hg), others classified concentration levels (low, medium, high). This heterogeneity prevented meta-analysis and makes direct comparisons of performance metrics (e.g., accuracy) tenuous.

Second, incomplete reporting of performance metrics. While most studies reported accuracy and AUC, many omitted precision, recall, F1-score, or specificity. This is problematic for pollution status classification, where class imbalance is common (e.g., contaminated samples are often rare). Without recall or F1-score, it is impossible to assess whether a model simply predicts the majority class (non-contaminated) and still achieves high accuracy. Only Qi et al. (2025) and Hu et al. (2024) explicitly addressed class imbalance using SMOTE or balanced random forest.

Third, lack of external validation. All 11 studies used internal cross-validation or a single held-out test set from the same source. None reported external validation on an independent dataset collected from a different site, instrument, or temporal period. This raises concerns about overfitting and limits the real-world applicability of the models, especially for spectroscopic and electrochemical sensors that are sensitive to matrix effects (pH, temperature, ionic strength, organic matter).

Fourth, limited matrix diversity and geographical bias. Water matrices dominate (72.7%), but even within water, most studies used spiked laboratory samples or controlled natural water (e.g., seawater, tap water) rather than naturally contaminated environmental waters with complex backgrounds. Soil studies were few (Hu et al., 2024; Qi et al., 2025), and sediment studies were absent. Geographically, 8 of 11 studies originated from China, the United States, and South Korea, raising questions about generalisability to other regions with different soil types, water chemistry, and regulatory frameworks.

Fifth, risk of bias in classification threshold justification. Our risk of bias assessment ( Figure 7, Table 12) revealed that several studies did not clearly justify the choice of regulatory threshold for defining “contaminated” vs. “non-contaminated” samples. For example, Valle et al. (2025) tested concentrations down to 1 nM for Hg but did not explicitly classify based on a regulatory limit, making it difficult to translate their findings into a binary pollution status decision.

4.2. Limitations of the review processes

This systematic review has several methodological limitations that should be acknowledged.

Search strategy and database coverage. The search was limited to three databases (Scopus, SpringerLink, ScienceDirect) and did not include Web of Science, PubMed, or IEEE Xplore. Although the combination of Scopus and SpringerLink provides broad coverage of environmental science and analytical chemistry literature, relevant studies in engineering, materials science, or biomedical sensors may have been missed. The final search date was 6 May 2026, so studies published after this date are not included.

Language and publication bias. Only peer-reviewed articles in English were included. This may exclude relevant studies published in other languages or non-peer-reviewed sources (e.g., conference proceedings, preprints). The exclusion of non-English studies may introduce language bias, particularly since heavy metal pollution is a global issue with substantial research output from non-English speaking countries.

Risk of bias assessment. Although we used a tailored tool adapted from JBI and QUADAS-2, the assessment was qualitative and subjective for some domains (e.g., handling of confounding factors). No meta-analysis was performed due to heterogeneity, so we could not statistically assess publication bias using funnel plots or Egger’s test.

Data extraction limitations. Some studies did not report complete performance metrics, and we recorded these as “not reported”. It is possible that contacting authors could have retrieved missing data, but this was beyond the scope of this review.

4.3. Implications for practice, policy, and future research

For environmental monitoring practice and regulatory agencies, current ML-based classification models have demonstrated high accuracy (up to 100%) in controlled settings, but practitioners should exercise caution when applying these models to new environmental matrices without revalidation. The absence of integrated spectroscopic-electrochemical approaches means that field-deployable sensors still rely on a single modality, potentially missing complementary information. Regulatory acceptance of ML-based methods will require standardised protocols for model training, validation, and reporting. Agencies such as the WHO and EPA would benefit from developing guidelines for ML-based pollution status classification, including minimum requirements for external validation and handling of class imbalance.

For future research, six priorities emerge from this review: (1) integration of spectroscopic and electrochemical data through data fusion strategies, which remains untested for pollution status classification; (2) expansion to underrepresented matrices (soil, sediment) and metals (Cr, Ni, Zn) using real-world contaminated samples; (3) external validation on independent datasets from different geographical locations, instruments, and time periods; (4) standardised reporting of accuracy, precision, recall, specificity, F1-score, and AUC, with balanced accuracy for imbalanced data; (5) explicit handling of class imbalance using SMOTE, RUSBoost, or BalancedRandomForest; and (6) incorporation of explainable AI (XAI) methods to enhance trust and understanding of ML decisions. Addressing these gaps will accelerate the translation of ML-based sensors from laboratory prototypes to field-deployable, regulatory-accepted monitoring systems.

5.1. Recommendation

Future research should prioritise the integration of spectroscopic and electrochemical data through data fusion strategies, as this remains the most critical gap identified in this review. No study to date has combined both modalities for pollution status classification, despite their complementary information. Additionally, researchers must expand to underrepresented matrices (soil and sediment) and heavy metals (Cr, Ni, Zn), using real-world contaminated samples rather than spiked laboratory solutions.

Methodologically, future studies should implement external validation on independent datasets from different geographical locations, standardise reporting of performance metrics (accuracy, recall, precision, F1-score, AUC), and explicitly address class imbalance using techniques such as SMOTE or BalancedRandomForest. Incorporating explainable AI methods will also enhance model interpretability and regulatory acceptance. Addressing these priorities will accelerate the translation of ML-based sensors from laboratory prototypes to field-deployable monitoring systems.