📄 medicine

Visibility Graph Analysis Identifies Informative Spectral Regions in Fiber Evanescent Wave Spectra of Synovial Fluid for Knee Osteoarthritis Diagnosis

This study demonstrates that applying visibility graph analysis to fiber evanescent wave spectra of synovial fluid effectively identifies specific informative spectral regions, particularly around 1636 cm⁻¹, thereby significantly improving the diagnostic accuracy for knee osteoarthritis compared to conventional machine learning methods applied to the full spectrum.

Original authors: Bohdan Mahlovanyi, Rafał Rak, Paweł Jakubczyk, Yaroslav Shpotyuk, Józef Cebulski, Catherine Boussard-Pledel, Bruno Bureau, Sławomir Snela, Karolina Jabłońska, Jarosław Jabłoński

Published 2026-09-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Bohdan Mahlovanyi, Rafał Rak, Paweł Jakubczyk, Yaroslav Shpotyuk, Józef Cebulski, Catherine Boussard-Pledel, Bruno Bureau, Sławomir Snela, Karolina Jabłońska, Jarosław Jabłoński

Original paper licensed under CC BY 4.0 (https://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

Knee osteoarthritis is a condition where the cushioning cartilage inside the joint slowly wears away, leading to pain and stiffness. For decades, doctors have relied on X-rays and magnetic resonance scans to see this damage, but these images only reveal the problem after the structure of the joint has already changed significantly. By the time a bone looks worn on a scan, the disease has often been progressing for years. Scientists have long sought a way to detect the disease much earlier, by looking for subtle chemical changes in the fluid that lubricates the joint. This fluid, known as synovial fluid, acts as a direct report card on the health of the knee, containing proteins and other molecules that shift as the joint begins to break down. The challenge has been that these chemical shifts are incredibly faint and hidden within a complex mixture, making them difficult to spot with standard laboratory tools.

A team of researchers recently tackled this problem by analyzing the synovial fluid of forty individuals: twenty patients with advanced knee osteoarthritis and twenty people without the disease. They used a technique called fiber evanescent wave spectroscopy, which involves shining infrared light through a special glass fiber dipped into the fluid. As the light passes through, different molecules absorb specific colors of the light, creating a unique spectral fingerprint for the sample. When the researchers applied standard computer algorithms to the entire spectrum of light absorption, the results were underwhelming. The best computer model could only distinguish between the sick and healthy groups about two-thirds of the time, a level of accuracy that is not good enough for a reliable medical test. This suggested that the disease signals were too weak to be found by looking at the whole picture at once.

To solve this, the researchers changed their approach entirely. Instead of treating the spectrum as a simple line of numbers, they transformed the data into a map of connections, a method known as visibility graph analysis. Imagine looking at a mountain range where the height of each peak represents the intensity of the light at that specific color. In this new method, the computer draws a straight line between any two points on the mountain range. If that line is not blocked by any higher ground in between, those two points are considered "visible" to each other. By counting how many other points each location can see, the computer creates a new profile that highlights the shape and structure of the data rather than just its height. This technique allowed the team to ignore the overall volume of the signal and focus on the unique patterns of the molecular arrangement.

When they applied this structural analysis, two specific regions of the light spectrum stood out as being significantly different between the two groups. The first was a band of light corresponding to the vibrations of protein bonds, and the second related to the vibrations of water and other hydrogen-rich molecules. The most powerful signal came from a narrow window centered at 1636 inverse centimeters, a range that captures the Amide I band, which is dominated by the stretching vibrations of the peptide bonds that hold proteins together. When the researchers fed only this specific slice of data into their classification models, the results improved dramatically. The computer could now correctly identify the disease with an accuracy that reached a score of 0.81, a substantial leap from the previous 0.67.

The findings suggest that the biochemical changes associated with knee osteoarthritis are not spread evenly across the entire chemical profile of the joint fluid. Instead, they are concentrated in specific molecular features, particularly those related to the structure of proteins. The disease appears to alter the way proteins are organized and how they interact with one another, creating a subtle but detectable shift in the way they absorb infrared light. While the differences are small, the new method proved capable of isolating them from the background noise of the fluid. The study indicates that by focusing on these specific structural patterns rather than the entire spectrum, it is possible to find a clearer signal of the disease.

However, the researchers are careful to note that this is a preliminary discovery. The study involved a relatively small group of patients, and the groups differed in age and gender, which could have influenced the results. The control group consisted of people undergoing surgery for knee injuries rather than perfectly healthy volunteers, which is a practical necessity but introduces variables that need to be managed in future work. The authors emphasize that while the method shows promise for identifying informative regions, it is not yet a ready-made diagnostic tool. Further testing with larger and more diverse groups is required to confirm that these spectral patterns are a reliable sign of osteoarthritis. For now, the work demonstrates that looking at the shape of the data, rather than just its size, can reveal hidden clues about the health of a joint.

Technical Summary: Visibility Graph Analysis for Knee Osteoarthritis Diagnosis via Synovial Fluid Spectroscopy

Problem Statement
Knee osteoarthritis (KOA) is a prevalent musculoskeletal disorder characterized by progressive cartilage degeneration and synovial inflammation. Current diagnostic methods, primarily radiography and MRI, detect structural changes only after significant disease progression. There is a need for molecular-level diagnostic tools capable of identifying biochemical alterations earlier. Synovial fluid, which reflects the local joint microenvironment, is a promising biological matrix for such analysis. However, distinguishing KOA from healthy controls using Fourier-transform infrared (FTIR) spectroscopy, specifically Fiber Evanescent Wave Spectroscopy (FEWS), is challenging. The biochemical differences between KOA and non-osteoarthritic synovial fluid are subtle, heterogeneous, and distributed across overlapping absorption bands. Conventional machine learning (ML) approaches applied to full-spectrum data often suffer from high dimensionality relative to sample size, leading to moderate classification performance and reduced interpretability.

Methodology
The study analyzed 40 synovial fluid samples (20 KOA patients and 20 controls) acquired via FEWS over the spectral range of 966–3750 cm⁻¹. The research employed a two-stage analytical framework:

Conventional Machine Learning Baseline: Nine supervised ML classifiers (including Logistic Regression with LASSO, SVM, Random Forests, and k-NN) were evaluated using 12-fold cross-validation on the full spectrum. The primary metric was the Area Under the Receiver Operating Characteristic Curve (ROC AUC).
Visibility Graph (VG) Analysis: To identify diagnostically relevant spectral regions without relying on absolute intensity, the authors transformed the 1D spectral signals into visibility graphs. In this topological approach, data points (wavenumbers) are converted into nodes, and edges are established based on a geometric "line-of-sight" criterion. The resulting graph topology is summarized by the "visibility degree" (node degree) for each point.
- The mean visibility degree profiles for the KOA and Control groups were compared to calculate the difference ( $\Delta v$ ).
- Regions with the largest differences in visibility degree were identified as potential informative spectral windows.
- Systematic searches were conducted across 682 window configurations (varying center positions and half-widths) using three classifiers (PLS-DA, KNN, XGBoost) to determine the optimal spectral interval for classification.

Key Results

Baseline Performance: Conventional ML applied to the full spectrum yielded only moderate discrimination. The best-performing model (Logistic Regression with LASSO) achieved a pooled ROC AUC of approximately 0.67, with a balanced accuracy of 0.675.
VG-Derived Insights: The visibility graph analysis revealed two distinct spectral regions exhibiting the largest structural differences between groups: 1500–1750 cm⁻¹ and 3150–3300 cm⁻¹.
Optimized Classification: Targeting the identified regions significantly improved performance. The best classification result was obtained using the PLS-DA classifier on a spectral window centered at 1636 cm⁻¹ with a width of 280 cm⁻¹ (spanning 1496–1776 cm⁻¹). This configuration achieved an ROC AUC of 0.81 ± 0.12 and a sensitivity of 1.0.
Spectral Interpretation: The optimal window (1496–1776 cm⁻¹) encompasses the Amide I (~~1647 cm⁻¹) and Amide II (~~1538 cm⁻¹) bands, as well as carbonyl stretching regions. These bands are associated with protein secondary structure, peptide bonds, and lipid/carboxyl groups, suggesting that KOA-related changes in synovial fluid are concentrated in protein conformation and composition.

Significance and Claims
The paper claims that visibility graph analysis serves as an effective, objective, and data-driven tool for identifying diagnostically relevant spectral intervals in complex biological spectra. By transforming spectral profiles into topological representations, the method highlights regions where local signal structure differs most between groups, independent of absolute intensity variations.

The authors state that this approach addresses the limitations of conventional full-spectrum ML by reducing the search space to biologically meaningful regions, thereby improving both the interpretability and predictive performance of spectroscopic classification. The study demonstrates that while KOA does not produce a single dominant spectroscopic biomarker, the disease manifests through subtle, systematic biochemical alterations concentrated in the Amide I/II domain. The integration of FEWS with VG analysis offers a novel pathway for investigating these molecular signatures, though the authors note that further validation in larger, independent cohorts is required before clinical application.

Limitations Noted by Authors
The study acknowledges several constraints: a relatively small sample size (n=40), potential confounding factors due to age and sex differences between the KOA and control groups, and the fact that the control group consisted of patients with acute trauma rather than asymptomatic healthy individuals. The authors emphasize that the identified spectral differences require validation in independent cohorts and correlation with established biochemical markers.

More like this