All research

Spectroscopic sensing · 2025–present

Generalisable deep learning for microbial SERS

Standardising surface-enhanced Raman spectra across instruments, recognising sepsis pathogens from blood, and examining how well microbial SERS models are validated.

Two research papers, one evidence review and ongoing work on spectral transfer

Graphical abstract showing spectra from portable and reference Raman instruments, device transformation, and bacterial classification.
Graphical abstract by Mahmud et al., Spectrochimica Acta Part A (2026). Reproduced under CC BY 4.0.View full-size figureSource articleCC BY 4.0

Approach

Surface-enhanced Raman spectroscopy (SERS) can detect microbial signatures quickly, but spectra change with the instrument, substrate, sample and acquisition conditions. Strong results can also be misleading when spectra from one biological sample appear in both training and testing. This project addresses those problems by standardising spectra across instruments, testing sepsis and pathogen recognition on blood-derived samples, auditing validation practice across the field, and extending spectral mapping to analytes unseen during training.

  1. Acquire and standardise spectra
  2. Recognise analytes and pathogens
  3. Test independence and transfer

Standardising SERS spectra across instruments

Spectrochimica Acta Part A · 2026 · Online 2025

Spectra from 20 bacterial and reference analyte classes were recorded on silver nanorod substrates with four portable Raman instruments and a laboratory Renishaw reference. SERS-D2DNet, a one-dimensional sequence-to-sequence network, learned to transform portable spectra toward the reference using 50 training spectra per class.

SuperRaman, a lightweight Super-ONN classifier, then identified the analyte classes before and after transformation. Agreement with the reference improved on all four instruments, and classification accuracy rose on each one. The combined framework uses 6.6 million parameters and takes 3.27 ms per inference.

R² with the reference after transformation
>98%
Enwave accuracy after transformation
99.98%
Analyte classes
20
R² against the Renishaw reference (Figure 6) and 20-class SuperRaman accuracy (Table 5), from Mahmud et al. (2026).
R² beforeR² afterAccuracy beforeAccuracy after
Enwave72.03%99.44%97.58%99.98%
Tec563.97%99.54%99.95%99.98%
Rapid ID81.70%98.80%96.54%98.87%
First Defender87.99%99.57%99.82%100.00%
SERS-D2DNet architecture with encoder and decoder blocks joined by full-scale skip connections, beside enlarged views of its multi-resolution blocks and feature layers.
Figure 2 from Mahmud et al. (2026). The SERS-D2DNet architecture and its main building blocks.View full-size figureSource articleCC BY 4.0
Bar chart of R² between transformed and reference spectra for four portable devices. SERS-D2DNet has the highest bar for every device, between 98.80% and 99.57%, above five benchmark models. The untransformed spectra score between 63.97% and 87.99%.
Figure 6 from Mahmud et al. (2026). R² for SERS-D2DNet and five benchmark models. The black bars compare untransformed portable spectra with the reference.View full-size figureSource articleCC BY 4.0
Bar chart of classification accuracy for four portable devices under four processing conditions. Accuracy rises with baseline correction and filtering, and domain transformation gives between 98.87% and 100%.
Figure 9 from Mahmud et al. (2026). Effect of baseline correction, filtering and domain transformation on classification accuracy.View full-size figureSource articleCC BY 4.0

Recognising sepsis and its causative pathogens from blood

Talanta · 2026 · Online 2025

The study collected SERS spectra from 653 analyte-enriched blood samples at a tertiary hospital in Qatar, comprising 271 sepsis-positive and 382 control samples. A separate external cohort added 70 samples. The tasks were binary sepsis recognition and six-class identification of control samples and five sepsis-causing bacteria.

SuperRamanNet is a compact one-dimensional Super-ONN classifier with about 0.65 million parameters. Cross-validation kept all spectra from each blood sample in the same fold, and augmentation was applied only to training data. Augmentation raised external pathogen accuracy from 66.67% to 98.28%, and saliency maps highlighted the 740–760 cm⁻¹ region across classes.

Sepsis recognition accuracy
99.67%
Six-class pathogen accuracy
98.84%
External pathogen accuracy
98.28%
SuperRamanNet results from Table 1 of Hassan et al. (2026).
AccuracyPrecisionRecallF1 scoreSpecificity
Sepsis recognition, five-fold99.67%99.64%100.00%99.82%95.50%
Pathogen identification, five-fold98.84%98.83%98.83%98.82%99.05%
Pathogen identification, external98.28%98.31%98.28%98.26%99.45%
Workflow illustration. Blood from sepsis and control subjects is centrifuged to enrich the analyte, which is placed on silver nanorod slides for SERS acquisition. The spectra are preprocessed and classified to separate control samples from five sepsis-causing bacterial strains.
Figure 1 from Hassan et al. (2026). Sample collection and analyte enrichment (a), followed by SERS acquisition, preprocessing and classification (b).View full-size figureSource articleCC BY 4.0
Two scatter plots, t-SNE and UMAP, of SuperRamanNet embeddings. Points form six separate clusters, one for each class: E. coli, K. pneumoniae, P. aeruginosa, S. aureus, S. epidermidis and negative controls.
Figure 5 from Hassan et al. (2026). t-SNE (a) and UMAP (b) projections of SuperRamanNet embeddings.View full-size figureSource articleCC BY 4.0
Six mean SERS spectra, one for controls and one for each of five bacteria, coloured by importance. Arrows mark the spectral bands that contributed most to each class.
Figure 6 from Hassan et al. (2026). Saliency maps showing the spectral regions that contributed most to each class.View full-size figureSource articleCC BY 4.0

Auditing validation across microbial SERS studies

Talanta · Online 2026

This PRISMA-guided review covered 129 experimental studies of label-free SERS with machine learning across bacterial, viral, fungal and cross-domain applications. It followed the full workflow, from target and matrix definition through sample handling, SERS platform design, spectral acquisition and preprocessing, modelling and interpretation, to validation.

The audit found that large spectral datasets often reflected repeated measurements from few biological samples. It concluded that deployable methods need biologically independent evaluation, leakage-safe preprocessing, reproducible substrates, calibration across instruments and matrices, and prospective multi-batch, multi-site validation.

Studies reviewed
129
Without biological-unit-aware validation
68.2%
With external or major-domain validation
6.2%
Corpus-level findings from the abstract of Mahmud et al. (2026).
Share of studies
Controlled-development datasets68.2%
No more than 25 independent biological units62.0%
No biological-unit-aware predictive validation68.2%
External or major-domain validation6.2%

PRISM-SERS: mapping spectra for unseen analytes

Ongoing research · Manuscript in preparation

PRISM-SERS (Platform-to-Reference, class-Independent Spectral Mapping) is my current work on a harder form of cross-device standardisation. It aims to translate portable-instrument spectra into the laboratory reference domain for analyte classes the model has never seen, using unpaired data and no class labels at inference.

Each analyte class is held out in turn, with its portable and reference spectra excluded from training, validation and model selection. The full evaluation is under way, and results will be reported once the multi-fold analysis is complete.

Spectra in the public dataset
29,545
Analyte classes held out in turn
20
Raman instruments
5

Research stages and outputs

  1. 2026 issue · Online 2025

    Standardising spectra across instruments

    Mapped spectra from four portable devices to a laboratory reference and tested how that affected classification across 20 analyte classes.

  2. 2026 issue · Online 2025

    Sepsis and pathogen recognition from blood spectra

    Developed SuperRamanNet for sepsis recognition and pathogen identification from analyte-enriched blood, with an independent external cohort.

  3. Online 2026

    Evaluating validation practices in the literature

    Reviewed 129 studies for biological independence, external evaluation and reporting across the workflow from sample preparation to modelling.

  4. Ongoing

    PRISM-SERS and transfer to unseen analytes

    Developing class-independent spectral mapping for analytes held out of training. A manuscript is in preparation, and a second workstream studies few-shot recognition.

My contribution

I led the cross-device study, including model development, experiments, generalisation analysis, visualisation and manuscript preparation. For the sepsis study, I developed the deep learning pipeline from raw spectra through preprocessing, data splitting, training, evaluation and interpretation. I also led the design, corpus audit and synthesis of the 129-study review, and I am developing PRISM-SERS.

Scope and validation

The cross-device study and PRISM-SERS use prepared analytes rather than clinical specimens. The sepsis study drew its primary samples from one hospital and tested a 70-sample external cohort, so prospective multi-site validation is still needed. The review percentages describe the 129 studies in its corpus, and PRISM-SERS has no published results yet.

Next steps

Complete the multi-fold PRISM-SERS evaluation and prepare the manuscript for submission.