🤖 machine learning

Machine Learning-Driven Multimodal Spectroscopic Liquid Biopsy for Early Multicancer Detection

This paper presents a machine learning-driven multimodal liquid biopsy framework that integrates FTIR, Raman, and EEM fluorescence spectroscopy to achieve highly accurate, label-free early detection of breast and colorectal cancers with ROC-AUC scores of 0.997 and 0.994, respectively.

Original authors: Alejandro Leonardo García Navarro, Javier Cachón Ortiz, Javier González Colsa, Samuel García Díaz, Carlos Viadero Valderrama

Published 2026-05-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Alejandro Leonardo García Navarro, Javier Cachón Ortiz, Javier González Colsa, Samuel García Díaz, Carlos Viadero Valderrama

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

Imagine you are trying to identify a specific type of fruit in a dark room. You could try to feel its shape (one sense), smell its scent (another sense), or listen to the sound it makes when you tap it (a third sense). While any one of these clues might help you guess the fruit, using all three together gives you a much clearer picture.

This paper is about doing exactly that, but instead of fruit, the researchers are trying to detect cancer (specifically breast and colorectal cancer) using blood samples. They call this a "liquid biopsy" because they are analyzing liquid (blood serum) instead of cutting out a piece of tissue.

Here is the simple breakdown of their "super-senses" and how they combined them:

The Three "Super-Senses"

The researchers used three different high-tech tools to look at the blood samples. Think of these as three different ways to "listen" to the chemical makeup of the blood:

FTIR Spectroscopy (The "Infrared Ear"): This tool shines infrared light at the blood. Different molecules (like proteins and fats) vibrate in specific ways when hit by this light. It's like listening to a choir where every singer (molecule) has a unique voice. This gives a broad overview of the blood's chemical "fingerprint."
Raman Spectroscopy (The "Laser Whisperer"): This uses a laser to bounce light off the molecules. It's very good at spotting tiny, specific details that the first tool might miss. If the first tool hears the whole choir, this one can hear exactly which violin is slightly out of tune.
EEM Fluorescence (The "Glow-in-the-Dark Detector"): This shines different colors of light at the sample and watches how it glows back. Some natural chemicals in the blood glow when hit by light. This tool maps out where these "glowing" chemicals are and how bright they are, revealing metabolic changes.

The Problem with Using Just One

The researchers tested each tool on its own.

The Result: Each tool was actually quite good at spotting cancer. It was like having a detective who is 95% sure of their answer.
The Catch: Sometimes Tool A was great at spotting breast cancer but missed colorectal cancer. Sometimes Tool B was the opposite. Relying on just one tool meant you might miss something or get a "false alarm."

The Solution: The "Team Huddle" (Multimodal Fusion)

Instead of letting each tool work alone, the researchers used Machine Learning (a type of computer brain) to combine all three tools into one super-team.

They didn't just ask the computer, "What did Tool A say?" They fed it the raw data from all three tools at the same time. They used a technique called Low-Level Data Fusion.

The Analogy: Imagine three experts looking at a crime scene. Expert A sees the footprints, Expert B sees the fingerprints, and Expert C sees the broken glass. If they talk to each other before making a conclusion, they can piece together a perfect story. If they work alone, they might miss the full picture.
The Process: The computer took the "footprints" (FTIR data), the "fingerprints" (Raman data), and the "broken glass" (EEM data), cleaned them up, and mashed them together into one giant puzzle. Then, a smart algorithm (called XGBoost) solved the puzzle to decide: "Is this cancer or is this healthy?"

The Results: The Power of Teamwork

When they tested this "Team Huddle" approach:

For Breast Cancer: The team got a score of 99.7% accuracy. This was slightly better than any single tool could do alone.
For Colorectal Cancer: The team got a score of 99.4% accuracy. Again, the combination was the most balanced and reliable.

The key finding wasn't just that the score went up a little bit; it was that the team was more consistent. Sometimes one tool was too sensitive (saying "cancer" when it wasn't), and sometimes another was too cautious. When they worked together, they balanced each other out, resulting in a diagnosis that was both highly accurate and very stable.

What They Did Not Claim

It is important to stick to what the paper actually says:

They did not say this is a cure for cancer.
They did not say this is ready to be used in every hospital tomorrow.
They did not claim they tested it on thousands of people from different countries yet.

The paper explicitly states that while the results are very promising, the group of people they tested on (about 300 people total) was still relatively small. They need to test this on much larger groups in the future to prove it works for everyone.

The Bottom Line

This paper shows that by combining three different ways of "listening" to blood chemistry and letting a computer brain analyze them all together, we can create a much more reliable early-warning system for cancer than using any single method alone. It's a proof-of-concept that says, "If we combine our senses, we can see the disease much clearer."

Technical Summary: Machine Learning-Driven Multimodal Spectroscopic Liquid Biopsy for Early Multicancer Detection

Problem Statement
Cancer remains a leading cause of global mortality, with projections indicating a significant rise in cases by 2050. Early detection is critical for improving survival rates and reducing the need for aggressive treatments. However, conventional diagnostic methods, such as tissue biopsies and medical imaging, face limitations including invasiveness, potential complications, tumor heterogeneity, and reduced sensitivity to early-stage biochemical alterations. While liquid biopsy offers a minimally invasive alternative, many current approaches rely on single biomarkers or single spectroscopic modalities, which may not capture the complex, systemic biochemical changes associated with cancer progression. There is a need for robust, label-free, and scalable diagnostic strategies that can holistically characterize biochemical alterations in biological fluids.

Methodology
The authors propose a multimodal spectroscopic liquid biopsy framework that integrates three distinct spectroscopic techniques with Machine Learning (ML) to detect breast and colorectal cancers.

Data Acquisition: Serum samples were collected from 100 breast cancer patients, 100 colorectal cancer patients, and 100 healthy controls. Three spectroscopic modalities were applied:
- Fourier Transform Infrared (FTIR) Spectroscopy: Performed on lyophilized serum using Attenuated Total Reflectance (ATR) in the 4000–650 cm⁻¹ range.
- Raman Spectroscopy: Performed on lyophilized serum using a 785 nm laser, capturing the 30–3358 cm⁻¹ range.
- Excitation–Emission Matrix (EEM) Fluorescence Spectroscopy: Performed on diluted serum (1:125 in PBS) using a spectrofluorometer, generating 2D matrices (250–520 nm excitation; 270–750 nm emission).
Preprocessing: Each modality underwent a dedicated preprocessing pipeline to remove artifacts (e.g., baseline drift, fluorescence background, scattering):
- FTIR: Evaluated 2,880 pipeline combinations; the selected workflow included polynomial baseline correction, Standard Normal Variate (SNV) scatter correction, Savitzky–Golay smoothing, and second derivative computation.
- Raman: Restricted to the 600–1800 cm⁻¹ region, with Asymmetric Least Squares (ALS) baseline correction and SNV scatter correction.
- EEM: Involved blank subtraction, physical masking (removing $\lambda_{em} < \lambda_{ex}$ ), and Rayleigh scatter removal.
Feature Fusion and Classification:
- Low-Level Data Fusion (LLDF): Feature vectors from the three modalities were concatenated. Prior to concatenation, features underwent z-score normalization and block scaling (dividing by $d^{1/4}$ , where $d$ is the feature dimension) to prevent dominance by high-dimensional modalities.
- Model: XGBoost classifiers were trained to distinguish cancer patients from healthy controls.
- Validation: Stratified 10-fold cross-validation was used. Patient-level grouping was enforced to prevent data leakage when replicates were present.
- Experimental Configurations: Seven configurations were evaluated: three unimodal (FTIR, Raman, EEM), three bimodal (FTIR+Raman, FTIR+EEM, Raman+EEM), and one full multimodal (FTIR+Raman+EEM).

Key Results
The study evaluated performance using ROC-AUC, balanced accuracy, sensitivity, and specificity.

Unimodal Performance: Individual modalities achieved high discrimination.
- Breast Cancer: Raman alone achieved the highest AUC (0.999). FTIR and EEM also performed strongly (AUC 0.979 and 0.982, respectively).
- Colorectal Cancer: Raman (AUC 0.983) and EEM (AUC 0.976) outperformed FTIR (AUC 0.961).
Multimodal Fusion: The integration of modalities generally improved robustness and balance, though optimal configurations varied by cancer type.
- Breast Cancer: The FTIR + Raman and Raman + EEM combinations both reached an AUC of 0.997. The full multimodal (FTIR + Raman + EEM) configuration achieved the highest sensitivity (0.990 ± 0.030) with an AUC of 0.997, though with slightly higher variability in specificity.
- Colorectal Cancer: The full multimodal configuration achieved the highest AUC (0.994 ± 0.019) with balanced sensitivity (0.959) and specificity (0.957). The FTIR + EEM bimodal configuration provided the most stable result with the highest balanced accuracy (0.961 ± 0.052) and specificity (0.963 ± 0.080).
Learning Curves: Analysis confirmed that the available cohort sizes were sufficient for stable model training, with performance plateauing rapidly.

Significance and Claims
The paper claims that the proposed framework demonstrates the feasibility of using multimodal spectroscopic liquid biopsy for early multicancer detection. The primary contribution is the demonstration that fusing complementary spectroscopic modalities (FTIR, Raman, and EEM) via low-level data fusion can enhance the robustness, sensitivity, and specificity of ML-based diagnostic models compared to unimodal approaches.

The authors emphasize that while individual modalities are effective, the multimodal approach provides a more balanced overall performance, particularly in colorectal cancer detection where the trimodal approach yielded the most consistent results. The study suggests that this approach leverages the complementary nature of the techniques: FTIR captures global molecular fingerprints, Raman provides specific vibrational information, and EEM detects endogenous fluorophores and metabolic alterations.

The authors maintain a modest tone regarding clinical deployment, acknowledging that the cohort size is moderate and that external validation on independent, multicentric cohorts is necessary to assess generalization. They position this work as a step toward developing robust, data-driven liquid biopsy platforms, rather than a finalized clinical tool. Future work is identified as focusing on larger studies, additional cancer types, and advanced fusion strategies.