Ewing sarcoma (ES) is the second most common malignant bone tumour in children and adolescents, comprising approximately 3% of all paediatric malignancies. It is a rare disease, affecting roughly 2.9 people per million annually. For patients with localised disease, overall survival approaches 69%, but patients who present with metastatic disease or who experience relapse face substantially worse odds: only about 42% survive five years. This disparity underscores how critical it is to identify, early in the treatment course, which patients are at high risk of treatment failure.
The case for risk-adapted therapy: Current ES treatment follows the Euro-EWING protocol, a multi-agent regimen that combines aggressive chemotherapy with surgery and, in some cases, radiation. Uniform treatment of all patients regardless of their biological risk profile means that some patients are overtreated (exposed to toxic regimens that yield no survival benefit) while others may receive insufficient therapy. Risk-adapted approaches, which have already improved outcomes in paediatric leukaemia and Hodgkin lymphoma, require reliable prognostic markers to stratify patients accurately before and during treatment.
Existing prognostic factors and their limitations: Established prognostic markers for ES include patient age, tumour size, tumour localisation, disease stage at diagnosis, and the degree of tumour necrosis observed in the surgically resected specimen after neoadjuvant chemotherapy. Of these, the percentage of tumour necrosis after neoadjuvant chemotherapy is considered the most clinically informative, with some researchers arguing that only patients achieving 100% necrosis should be classified as good responders. However, necrosis assessment is available only after surgery, which occurs months into the treatment course. A tool that could predict treatment outcome before or early in chemotherapy would have far greater clinical utility.
This 2019 study, published in Molecules (MDPI), presents a proof-of-concept investigation into whether Fourier transform infrared (FTIR) spectroscopy of bone biopsy tissue, combined with machine learning classifiers, can predict ES treatment outcomes, including patient death, disease relapse, and chemotherapy response, with clinically meaningful accuracy.
Fourier transform infrared (FTIR) spectroscopy is a physicochemical technique that probes the bulk biochemical composition of a biological sample by measuring how different molecular bonds absorb infrared radiation at characteristic frequencies. Because the frequency range at which a molecule absorbs infrared light depends on its chemical structure, FTIR spectra function as a molecular fingerprint, reflecting the relative abundance of proteins, nucleic acids, lipids, polysaccharides, and phospholipids within a tissue section. The technique is non-destructive and label-free, requiring no staining or chemical modification of the tissue.
Application to cancer tissue: FTIR has been applied across a wide range of cancer types, including breast, lung, ovarian, brain, cervical, gastric, colon, and prostate cancers, to detect compositional changes that distinguish cancerous from healthy tissue or that differentiate tumour subtypes. The spectral region of 800-3500 cm-1 encompasses the vibrational frequencies of virtually all biologically relevant molecular classes, while the "bio-fingerprinting" region of 1000-1100 cm-1 is particularly informative for nucleic acids and polysaccharides. Earlier work by the same research group had demonstrated that peak absorbance maxima in this fingerprinting region could predict ES patient outcomes, establishing the conceptual basis for this expanded study.
ATR-FTIR setup used in this study: Spectra were recorded using a Bruker Vertex 70v FTIR spectrometer. Tissue specimens were applied to an attenuated total reflection (ATR) plate and mid-infrared radiation was passed through the sample using a single-reflection snap ATR crystal diamond, with measurements recorded at 1 cm-1 spectral resolution over 32 scans per specimen in the range of 800-3500 cm-1. All measurements were performed in triplicate to reduce measurement noise. Formalin-fixed paraffin-embedded (FFPE) tissue sections from biopsy samples were used, sectioned to 10 micrometres thickness on calcium fluoride slides and deparaffinised before measurement.
The key innovation of this study compared to prior work is the shift from analysing specific peak features to feeding the entire FTIR spectrum, spanning thousands of individual absorbance values across the measured wavenumber range, into machine learning classifiers. This assumption-free, data-driven approach allows the classifier to identify spectral patterns associated with clinical outcomes that may not correspond to any single identifiable peak or functional group.
Twenty-seven patients aged 5 to 20 years with newly diagnosed Ewing sarcoma of bone were enrolled in this retrospective study. Patients were treated at the Department of Surgical Oncology, Institute of Mother and Child, Warsaw, Poland, according to Euro-EWING protocols between 2010 and 2016. Every patient received identical induction neoadjuvant chemotherapy (neoCTX) consisting of six VIDE cycles, a four-drug regimen of vincristine, ifosfamide, doxorubicin, and etoposide. Microscopically complete resection was achievable in 26 of 27 patients. Histopathological assessment of chemotherapy response was measured as the percentage of viable tumour cells remaining after neoCTX completion, with a good response defined as at least 90% tumour necrosis.
Tissue samples and time points: FFPE bone tissue blocks were collected from two time points for each patient: a diagnostic biopsy obtained before neoCTX began (preCTX), and a surgical resection specimen collected after the sixth VIDE cycle was completed (postCTX). This paired design allows comparison of the tumour's biochemical composition before and after chemotherapy, and the spectral change between the two time points (the difference spectrum, preCTX minus postCTX) was also analysed as a third feature set. All samples were prepared and verified by pathologists with specific expertise in ES.
Outcome variables: Three distinct clinical outcomes were modelled as binary classification targets: patient death due to cancer progression, disease relapse, and achievement of high tumour necrosis (greater than 90%) after neoCTX. These represent clinically distinct and increasingly challenging prediction tasks. Mortality and relapse are direct survival endpoints, while necrosis percentage is a surrogate marker of chemotherapy sensitivity that is used clinically to guide post-operative treatment decisions.
Cross-validation strategy: Given the cohort size of 27 patients, use of a held-out external test set was not feasible. Leave-one-out cross-validation (LOOCV) was used throughout, meaning each patient was held out once as the test case while the model was trained on the remaining 26 patients, cycling through all 27 patients. LOOCV provides an approximately unbiased estimate of generalisation accuracy for small datasets, though it does not substitute for true external validation on an independent cohort.
Raw FTIR spectra are high-dimensional objects: each spectrum consists of absorbance values at thousands of individual wavenumbers across the 800-3500 cm-1 range. Feeding this raw, high-dimensional input directly into a classifier would lead to overfitting, particularly with only 27 training samples. The preprocessing pipeline therefore applied mean-centering, scaling to unit variance (StandardScaler from Scikit-Learn), smoothing via a Savitzky-Golay filter, and a linear detrend to remove baseline drift. This standardised the spectra and reduced noise before dimensionality reduction.
Unsupervised dimensionality reduction for exploratory analysis: The first step was to assess whether FTIR spectra could visually separate patients by clinical outcome without any supervised labels. Standard principal component analysis (PCA) showed some clustering of patients who lived versus those who died in the first two principal components, but with considerable overlap. A suite of additional non-linear and matrix decomposition methods was also tested, including kernel PCA variants (using sigmoid, polynomial, cosine, and radial basis function kernels), as well as manifold learning approaches such as t-SNE and UMAP. Kernel PCA with a cosine function produced the clearest visual separation of outcome classes, particularly for mortality, motivating its use in the supervised modelling pipeline.
Feature spaces for supervised classification: Four distinct feature set configurations were generated. The first used only preCTX spectra, reduced to 25 dimensions using PCA (15 components) and kernel PCA with cosine function (10 components). The second used only postCTX spectra, similarly reduced to 25 dimensions. The third used the spectral difference (preCTX minus postCTX) as the feature set. The fourth concatenated the dimensionally reduced preCTX and postCTX representations into a 50-dimensional combined feature space. These four configurations were systematically evaluated across all three clinical outcome endpoints and all classifier architectures.
Dimensionality reduction via PCA and kernel PCA served two purposes: it reduced the contribution of spectral noise to the classification task, and it collapsed the thousands of correlated wavenumber variables into a manageable number of orthogonal features that could be handled by small-sample classifiers without overfitting. The inclusion of both linear (PCA) and non-linear (kernel PCA) components reflected the earlier finding that non-linear transformations better separated clinical outcome classes in the exploratory analysis.
Five classifier families were benchmarked on the four feature set configurations using LOOCV: linear Support Vector Machine (SVM), Random Forest (RF), Decision Tree, Linear Discriminant Analysis (LDA), and Gradient Boosted Classifier (GBM). All classifiers were used with their default hyperparameters as implemented in Scikit-Learn, a design choice that limits the risk of hyperparameter overfitting to the small training set. The classification accuracies reported are LOOCV percentages across 27 patients.
Mortality prediction: For predicting patient death due to cancer progression, LDA trained on preCTX spectra alone achieved 81% accuracy. When postCTX spectra were used instead, a Random Forest classifier reached better than 92% accuracy. Combining the preCTX and postCTX feature sets (concatenated 50-dimensional space) yielded an LDA model with 92% accuracy, corresponding to only two misclassified patients. The spectral difference feature set generally underperformed for mortality prediction, except with the k-nearest neighbours (KNN) classifier. A key additional finding emerged when only absorbances below 1800 cm-1 were analysed (excluding the lipid-dominated region that can be altered by deparaffinisation): a linear SVM using this restricted spectral range achieved 96.2% accuracy for mortality prediction, corresponding to just one misclassified patient.
Relapse prediction: Relapse prediction was described as more challenging than mortality prediction. The best-performing model was a linear SVM trained on preCTX spectra alone, achieving 92% accuracy with only two misclassifications. Concatenating or subtracting the preCTX and postCTX spectra did not improve relapse prediction accuracy, unlike what was observed for mortality prediction. This suggests that the biochemical state of the tumour at diagnosis, before chemotherapy is administered, carries substantial predictive information for relapse risk, an important finding because it implies that a pre-treatment biopsy alone could provide prognostic information.
Necrosis prediction: Predicting whether a tumour would achieve greater than 90% necrosis after chemotherapy proved most difficult. KNN and linear SVM trained on preCTX spectra achieved only 77% accuracy. Using the spectral difference (postCTX minus preCTX) as features improved linear SVM accuracy to 84.6% for necrosis prediction, consistent with the intuition that treatment-induced compositional changes carry the most information about chemotherapy sensitivity. This 84.6% accuracy is also consistent with results from Bergner et al., who found that distinguishing necrotic from tumorous tissue by FTIR spectral imaging achieved sensitivity of 75.3%. Across all three clinical endpoints, linear SVM classifiers delivered the most consistent performance.
To assess whether the classifier predictions were capturing genuine prognostic signal rather than statistical noise, the authors conducted formal survival analysis using Kaplan-Meier curves and Cox proportional hazard models. The median follow-up for the 27 patients was 29 months, with a range of 14 to 74 months. All deaths in the study cohort were attributable to cancer progression; there were no deaths from treatment toxicity or unrelated causes. The three-year progression-free survival rate for the cohort was 41.36%, and the three-year overall survival rate was 56.66%, figures consistent with the known prognosis of paediatric ES in a mixed localised-and-metastatic cohort.
Statistical significance of predictions: Kaplan-Meier plots were stratified by LOOCV model predictions (predicted to die versus predicted to survive, and predicted to relapse versus predicted not to relapse). Log-rank tests compared survival curves between model-predicted groups. All three primary models demonstrated highly statistically significant separation: the LDA mortality model using combined preCTX+postCTX features achieved a log-rank p-value of 0.00023; the Random Forest postCTX mortality model achieved p = 0.0002; and the linear SVM relapse model achieved p = 0.0004. These p-values indicate that the spectral models are capturing real differences in patient biology rather than overfitting noise.
Predictive accuracy in longitudinal context: In the survival analysis, the authors noted that in all but one case, patients predicted to die by either the LDA or RF mortality models did in fact die during the follow-up period. This observation, while based on a small cohort, suggests high positive predictive value for the mortality classifiers in practice. Cox proportional hazard models confirmed that the predicted risk group (model-positive versus model-negative) was a statistically significant predictor of survival time, even after accounting for the small sample size.
The strong statistical significance of these survival curve separations provides an important layer of validation beyond classification accuracy alone. A classifier could achieve high LOOCV accuracy through imbalanced class distributions, but the log-rank p-values below 0.001 for each primary model suggest the predictions are tracking genuine differences in survival outcome, not simply predicting the majority class.
Small single-centre cohort: The most significant limitation of this study is the cohort size of 27 patients from a single institution. Ewing sarcoma is a rare disease by definition, and even at a specialist centre, accruing large numbers of matched pre- and post-treatment specimens is exceptionally difficult. With 27 samples, leave-one-out cross-validation was the only statistically defensible validation strategy, but LOOCV accuracy estimates are known to have high variance and can overestimate generalisation performance. The authors acknowledge this explicitly, noting that the results should be interpreted carefully and that confirmation will require larger sample sizes.
Absence of external validation: No independent external test set was used to validate the models. All reported accuracy figures are LOOCV estimates on the same 27-patient cohort used for training. Performance on an unseen cohort from a different institution, with different tissue processing protocols, different spectrometer hardware, or a different patient population, remains entirely unknown. Machine learning models trained on single-centre data, particularly with small samples, frequently show performance degradation on external validation, sometimes of 10-15% or more.
Deparaffinisation and lipid-region uncertainty: The FFPE preparation and deparaffinisation process may alter lipid-associated spectral features in the 2800-3000 cm-1 range. The authors addressed this concern directly by repeating analyses restricted to absorbances below 1800 cm-1, excluding the lipid region. Reassuringly, the average change in classification accuracy across all models was less than 1% when the lipid region was excluded, suggesting that the predictive signal is primarily located in non-lipid spectral features. Interestingly, the sub-1800 cm-1 linear SVM model for mortality achieved an even higher accuracy of 96.2% compared to the full-spectrum model, which the authors note corresponds to only one misclassified patient.
Treatment homogeneity and generalisability: All 27 patients received identical VIDE-based neoadjuvant chemotherapy. The authors explicitly note that the results should be considered applicable only to VIDE-treated patients and cannot be assumed to generalise to other chemotherapy regimens. Additionally, the cohort was restricted to patients with complete clinical data and paired specimens from both biopsy and surgical resection, which may introduce selection bias toward patients who completed the full treatment protocol.
Scaling up to clinically viable sample sizes: The authors state explicitly that this study was designed not to produce a clinically deployable predictive model, but to validate the potential of FTIR spectroscopy as a prognostic tool. A clinically applicable model would require a larger multicentre trial with hundreds of patients to enable robust training, a held-out external validation cohort, and ideally prospective collection of specimens according to a standardised protocol. Given that ES affects approximately 2.9 people per million annually, such a study would likely require international collaboration across multiple specialist paediatric oncology centres, potentially through existing consortia such as the European Ewing Tumour Working Initiative of National Groups (EURO-E.W.I.N.G.).
Extending to other chemotherapy regimens: Current results are restricted to patients treated with VIDE induction. Euro-EWING protocols have evolved over time, and other regions use different regimens. Determining whether FTIR-based predictive models generalise across different induction chemotherapy backbones would be important for broad clinical application. Cross-validation of models across cohorts treated with VDC/IE (vincristine, doxorubicin, cyclophosphamide alternating with ifosfamide and etoposide) or other regimens would be necessary.
Mechanistic interpretation of spectral changes: An important open question is which specific biochemical changes in the tumour tissue drive the predictive power of the FTIR spectra. The exploratory analysis showed that differences in lipid and nucleic acid spectral intensity were observed between patients who lived and those who died after chemotherapy, with an increase in protein amide signal in those who died. Deeper investigation of these spectral differences, potentially combined with parallel RNA sequencing, proteomics, or metabolomics on the same tissue samples, could provide mechanistic insight into the biology of chemotherapy resistance in ES.
Integration with existing clinical prognostic factors: Future models could combine FTIR spectral features with established clinical prognostic variables (tumour size, stage, patient age, presence of metastases) and emerging molecular biomarkers (circulating tumour DNA, biomarkers such as LGALS3BP and nucleophosmin that have been investigated in ES). A multimodal prognostic model integrating FTIR spectra with clinical and molecular data might achieve greater discriminative power than any single data type alone, while also providing a richer biological picture of individual patient risk.