Chondrosarcoma is the second most common primary bone malignancy after osteosarcoma, accounting for approximately 20-27% of all primary bone tumors. Unlike many other cancers, chondrosarcoma shows a conspicuous resistance to chemotherapy and radiation, making surgical resection the only curative treatment option. The extent of resection, and therefore the prognosis, depends heavily on the accuracy of preoperative grading: grade I chondrosarcomas (also called atypical cartilaginous tumors or ACT in peripheral locations) can be managed conservatively with curettage, while grades II and III require wide surgical excision with larger functional sacrifice.
The grading dilemma: Distinguishing between benign enchondromas, atypical cartilaginous tumors (ACT/grade I), and true grade II-III chondrosarcomas represents one of the most contentious diagnostic problems in musculoskeletal oncology. MRI is the current workhorse of preoperative evaluation, but it cannot reliably differentiate ACT from enchondroma or accurately grade chondrosarcoma. Biopsy, the gold standard, carries risks of tract contamination, sampling error, and procedural morbidity, and even histopathologic assessment is subject to significant interobserver variability between pathologists.
Why exhaled breath: Tumor cells alter systemic metabolism in ways that produce distinctive volatile organic compounds (VOCs), low-molecular-weight chemicals that diffuse into the bloodstream from tissues and are subsequently exhaled through the lungs. The concept of VOC breath analysis as a liquid biopsy surrogate has been explored in lung, colorectal, and breast cancers. The hypothesis applied here is that chondrosarcoma, a tumor of chondrocyte origin with specific metabolic alterations including hypoxia-driven anaerobic pathways, generates a detectable VOC signature distinguishable from that of healthy individuals and benign cartilaginous lesions.
This 2024 pilot study from Leiden University Medical Center, published in Future Oncology, is one of the first to directly test whether an electronic nose (the aeoNose device) paired with an artificial neural network classifier can separate chondrosarcoma patients from healthy controls and patients with benign bone lesions based on their exhaled VOC profiles.
The study enrolled 57 participants between 2018 and 2023 at a single tertiary sarcoma referral center (Leiden University Medical Center, the Netherlands). The cohort comprised three groups: 24 patients with histologically confirmed chondrosarcoma, 25 healthy controls without known malignancy or inflammatory disease, and 8 patients with benign cartilaginous or bone lesions including enchondroma and other histologically confirmed benign entities. This case-control design is standard for early-phase diagnostic pilot studies but introduces important caveats regarding spectrum bias discussed in the limitations section.
The aeoNose device: The aeoNose (the eNose Company, Zutphen, the Netherlands) is a handheld electronic nose containing an array of metal-oxide semiconductor (MOS) sensors. Each sensor in the array responds differently to different chemical classes of VOCs, producing a resistance-change pattern that serves as a chemical fingerprint of the breath sample. The patient breathes through the device for a defined collection period, during which the sensor array records the evolving resistance signal. Unlike mass spectrometry-based VOC profiling, the aeoNose does not identify specific chemical compounds; instead, it captures a multivariate sensor pattern that encodes the overall VOC profile. This approach sacrifices chemical specificity for the sake of simplicity, portability, and speed.
Artificial neural network classifier: The sensor patterns from all participants were used to train an artificial neural network (ANN) to discriminate between the diagnostic groups. The ANN architecture consisted of input nodes corresponding to the sensor features extracted from the aeoNose signal, one or more hidden layers with nonlinear activation functions, and an output classification layer. Two models were developed: Model 1 compared chondrosarcoma patients directly against healthy controls (the primary binary classification task), and Model 2 incorporated a broader comparison potentially including benign lesion patients to assess robustness. Model performance was evaluated using receiver operating characteristic (ROC) analysis and cross-validation to account for the limited sample size.
Breath collection protocol: Standardized breath collection protocols were employed to minimize confounding. Patients were instructed to avoid eating, smoking, and strenuous exercise for a defined period prior to testing. All samples were collected in a controlled clinical environment. Despite these precautions, residual environmental and lifestyle confounders affecting exhaled VOC composition (including ambient air quality, medications, and comorbidities) represent an inherent challenge in breath-based diagnostics, particularly in a small pilot cohort.
The primary performance metric for Model 1 (chondrosarcoma versus healthy controls) was an area under the receiver operating characteristic curve (AUC) of 0.66. This result sits in the "poor to fair" range of diagnostic discrimination, where AUC = 0.5 represents chance classification and AUC = 1.0 represents perfect discrimination. At the operating threshold selected to balance sensitivity and specificity, Model 1 achieved a sensitivity of 75% and a specificity of 65%. In practical terms, this means the model correctly identified 75% of true chondrosarcoma cases as positive, while correctly ruling out disease in 65% of healthy individuals. The false positive rate of 35% is particularly notable from a clinical standpoint, as it would generate a substantial number of unnecessary further investigations in a screening context.
Model 2 results: The second model, which incorporated a modified comparison group, achieved a marginally improved AUC of 0.69. While this represents a modest gain over Model 1, both figures remain well below the thresholds typically considered acceptable for standalone clinical diagnostic tools (generally AUC greater than 0.80 for a reasonable rule-in test, and AUC greater than 0.90 for a high-confidence rule-out test). The scatter plots of cross-validated predicted values illustrate the substantial overlap between chondrosarcoma patient scores and healthy control scores, visually confirming the limited separation capacity of the current model.
Interpreting the ROC curves: The ROC curves presented for both models show a curve that hugs close to the diagonal line of no discrimination, rising only modestly above it. This is a qualitatively different performance profile from high-performing AI diagnostic tools described in imaging or pathology literature, where ROC curves often sweep to the upper left corner of the plot with AUC values of 0.85 to 0.98. The modest performance here is not surprising given the small cohort, the heterogeneity within the chondrosarcoma group (which likely includes grade I, II, and III tumors with different metabolic activity levels), and the known challenge that metal-oxide sensor arrays lack the chemical resolution of mass spectrometry-based VOC profiling.
No subgroup analysis by grade: A notable gap in the results is the absence of a stratified analysis comparing the classifier's performance across chondrosarcoma grades. Since the primary clinical question in chondrosarcoma management is not merely "cancer vs. no cancer" but "low-grade vs. high-grade vs. benign," understanding whether the VOC signature correlates with grade would be a critical secondary finding. The pilot study's sample size likely precluded such stratification, but it represents a key analytical objective for future work.
The field of breath-based diagnostics rests on the observation that exhaled air contains hundreds to thousands of VOCs, many of which reflect the biochemical state of the body. Healthy cells, inflamed tissues, and malignant tumors each produce different metabolic byproducts that volatilize into the blood and are cleared through pulmonary gas exchange. In cancer, increased glycolysis (the Warburg effect), lipid peroxidation from oxidative stress, and tumor-specific enzymatic activities generate VOCs including alkanes, aldehydes, ketones, and aromatic compounds at concentrations in the parts-per-billion to parts-per-trillion range.
How metal-oxide sensors work: Metal-oxide semiconductor (MOS) sensors function by detecting changes in electrical resistance when VOC molecules adsorb onto a heated metal-oxide surface. Different metal oxides (such as tin dioxide, zinc oxide, and tungsten trioxide) respond preferentially to different chemical classes. An array of MOS sensors with different materials thus produces a multivariate resistance-change pattern that encodes information about the overall VOC composition of the breath sample, even without identifying specific molecules. This is analogous to how a sommelier's palate integrates thousands of molecular stimuli into a holistic flavor assessment without necessarily identifying every individual compound.
Artificial neural network architecture: The ANN classifier used in this study functions as a nonlinear pattern recognition system. The input layer receives the time-series sensor resistance signals (or summary features extracted from them), passes these through hidden layers where weighted connections and activation functions transform the data into increasingly abstract representations, and produces a final output score indicating the probability of chondrosarcoma. Unlike simpler classifiers such as logistic regression or support vector machines, neural networks can model complex nonlinear interactions between sensor features, which is theoretically advantageous for the complex VOC mixture patterns generated by breath samples. However, this added complexity also increases the risk of overfitting when training data are limited, a concern directly relevant to this 57-patient cohort.
Comparison with mass spectrometry approaches: The aeoNose approach trades chemical specificity for clinical practicality. Gas chromatography-mass spectrometry (GC-MS) and proton transfer reaction-mass spectrometry (PTR-MS) can identify and quantify individual VOC compounds in exhaled breath at very low concentrations and have been used in research to identify specific cancer-associated VOC biomarkers. However, these techniques require specialized laboratory infrastructure, lengthy analysis times, and trained operators. The aeoNose by contrast is a portable point-of-care device that provides a result within minutes. The trade-off is that the sensor pattern it generates captures less chemical information and cannot identify which specific VOC compounds drive the classification signal, making mechanistic interpretation difficult.
The current diagnostic workup for suspected chondrosarcoma typically begins with conventional radiography revealing a calcified intramedullary lesion with endosteal scalloping or cortical expansion, followed by MRI to characterize lesion morphology, soft tissue extension, and degree of endosteal involvement. Cross-sectional imaging with CT provides additional information about matrix mineralization pattern. When imaging features are indeterminate or suggest higher-grade tumor, a percutaneous CT-guided biopsy or incisional biopsy is performed, with all the attendant procedural risks and sampling challenges described above.
The questionable cases: The most clinically difficult scenario in chondrosarcoma diagnosis is the incidentally discovered or symptomatic cartilaginous lesion in which MRI cannot confidently distinguish an enchondroma from an ACT/grade I chondrosarcoma. These cases consume significant resources in multidisciplinary sarcoma tumor board discussions, repeated imaging, and often ultimately proceed to biopsy or even surgery despite uncertainty. A reliable noninvasive biomarker that could preoperatively stratify these cases would have direct clinical value by potentially reducing the number of biopsies performed and informing the surgical approach when intervention is needed.
Adjunctive rather than standalone role: The authors themselves conclude that based on the current results (AUC 0.66-0.69), the aeoNose cannot be recommended as a standalone diagnostic biomarker for chondrosarcoma in daily practice. However, they propose that it might serve an adjunctive role alongside MRI in cases where imaging findings are borderline. This is a reasonable framing for a first-generation technology with limited discriminative performance: not as a replacement for existing workup, but as an additional data point that could shift pretest probability in ambiguous cases. The key question is whether the modest classification accuracy demonstrated here would translate into meaningful clinical decision support when applied on top of MRI in the specific subset of "questionable" cases.
Precedent from other cancers: The use of electronic nose technology has been investigated more extensively in lung cancer, where the proximity of the tumor to the airways provides a more direct route for VOC shedding into exhaled breath. Studies in lung cancer have reported considerably higher AUC values (0.80-0.95) for electronic nose classifiers, though these also involve larger cohorts and more homogeneous patient populations. The lower performance in chondrosarcoma likely reflects the greater physical distance between the bone tumor and the pulmonary gas exchange surface, the lower metabolic activity of many chondrosarcomas compared to high-grade carcinomas, and the smaller training sample.
Small cohort and statistical power: With 24 chondrosarcoma patients and 25 healthy controls, this is an underpowered exploratory study. Training an artificial neural network on a dataset this size introduces substantial risk of overfitting, where the model learns idiosyncratic features of the training individuals rather than generalizable biological signal. The use of cross-validation mitigates but does not eliminate this risk. Published guidance on minimum sample sizes for AI-based diagnostic model development generally recommends at least 10 events per input feature; the small number of participants relative to the multidimensional sensor input space means the model is likely operating at the boundary of statistical validity.
Spectrum bias: The comparison groups in this study do not fully represent the clinical decision problem. Comparing chondrosarcoma patients to healthy volunteers creates an artificially wide diagnostic contrast. In real clinical practice, the relevant comparison is between chondrosarcoma and patients with benign enchondromas or ACTs presenting with similar imaging features and symptoms. The 8-patient benign lesion group is too small to support meaningful subgroup analysis against this more clinically relevant reference population. A study using "questionable cartilaginous lesions" as the control group, rather than healthy controls, would provide a more clinically informative performance estimate.
Heterogeneity within the chondrosarcoma group: Chondrosarcoma encompasses a range of histologic grades (I, II, III), anatomical locations (central vs. peripheral vs. dedifferentiated), and molecular subtypes, each with different metabolic profiles and VOC-generating capacity. Pooling all chondrosarcoma grades into a single case group assumes a uniform metabolic signature across this heterogeneous entity, which is unlikely to hold. Grade I tumors (ACT) are low-grade, slow-growing, and may have minimal metabolic derangement compared to healthy cartilage, while grade III tumors have high proliferative rates and active Warburg metabolism. This heterogeneity within the cancer group would be expected to reduce classifier discriminative power.
Confounders and standardization: Exhaled VOC composition is affected by numerous factors beyond tumor biology, including comorbidities (diabetes, chronic lung disease, liver disease), medications, smoking history, diet, time of day, and ambient air quality. In a 57-person cohort, even well-intentioned standardization protocols cannot fully control for these variables. The authors do not report detailed baseline characteristics comparing chondrosarcoma patients and healthy controls for these confounders, which limits interpretation of whether the classification signal reflects tumor-specific VOC biology or group-level demographic and lifestyle differences.
The application of AI to sarcoma diagnostics is a rapidly developing field that has progressed along several parallel tracks. Radiomics-based approaches applying convolutional neural networks and random forest classifiers to MRI and CT features have demonstrated AUC values of 0.75-0.92 for distinguishing chondrosarcoma grades or differentiating malignant from benign bone tumors, with studies such as those using T1, T2, and contrast-enhanced MRI sequences showing particular promise for the enchondroma vs. ACT discrimination problem. Deep learning models applied to digitized histopathology slides of cartilaginous tumors have shown high accuracy in distinguishing enchondroma from chondrosarcoma and grading malignant tumors, with some studies reporting accuracy exceeding 85% for grade classification tasks.
Liquid biopsy and molecular approaches: Circulating tumor DNA (ctDNA) and cell-free DNA (cfDNA) methylation profiling are emerging noninvasive approaches for sarcoma diagnosis that offer higher molecular specificity than VOC breath analysis. IDH1/IDH2 mutations, which are present in 50-70% of central chondrosarcomas, can in principle be detected in liquid biopsy samples and provide both diagnostic confirmation and potential therapeutic relevance. Similarly, cfDNA methylation signatures specific to chondrosarcoma have been identified in small profiling studies. These approaches have higher potential diagnostic specificity than breath VOC analysis but require specialized laboratory infrastructure and blood sampling.
Complementary roles: The VOC breath analysis approach studied here occupies a distinct niche among noninvasive diagnostic modalities. Its primary advantage is the complete absence of any biological sampling, making it genuinely noninvasive, point-of-care, and repeatable without patient burden. If VOC breath analysis could be improved to achieve acceptable discriminative accuracy (AUC greater than 0.80), it could serve as a first-line triage tool in the diagnostic algorithm, identifying patients who warrant expedited imaging and biopsy while reassuring those with low probability of malignancy. Integration with MRI-derived radiomic scores into a multimodal AI classifier is one pathway toward improved performance that future studies could explore.
The current study's honest reporting of modest results is itself valuable for the field. In an era where research publication bias strongly favors positive findings, pilot studies that demonstrate marginal performance of a promising technology provide important calibration data and set realistic expectations for the sample sizes and design improvements needed before clinical adoption is warranted.
Sample size and multicenter design: A properly powered validation study for an electronic nose-based chondrosarcoma classifier would require substantially larger cohorts, likely 200 or more cases per group, assembled across multiple sarcoma referral centers to ensure the model generalizes across different patient populations, comorbidity profiles, and geographic VOC background compositions. Multicenter designs also provide the statistical power to perform stratified analyses by tumor grade, location (central vs. peripheral), and IDH mutation status, which would be essential for understanding the biology driving any classification signal and the subset of patients in whom the test performs best.
Clinically relevant control groups: Future studies should compare chondrosarcoma patients against patients with enchondromas, ACT/grade I tumors, and other benign bone lesions presenting with similar clinical and imaging features rather than healthy volunteers. This would test the technology against the actual clinical diagnostic challenge and generate sensitivity and specificity estimates that are meaningful for clinical decision-making rather than inflated by the easy comparison against symptom-free healthy controls.
Advanced sensor technologies and compound identification: The limited discriminative performance of metal-oxide sensor arrays may be partly overcome by using ion mobility spectrometry or compact mass spectrometry modules that can identify specific VOC compounds in exhaled breath. Knowing which specific compounds drive the chondrosarcoma signal would enable targeted optimization of sensor arrays for those compounds, potentially improving sensitivity and specificity. Studies using GC-MS to profile chondrosarcoma exhaled breath and identify candidate biomarker compounds would be a valuable precursor to optimized electronic nose development.
Integration with imaging AI: A multimodal diagnostic model combining the aeoNose output with MRI radiomic features and clinical variables (patient age, tumor location, imaging grade) into a unified prediction score could potentially outperform either modality alone. This integration approach is analogous to multi-modal AI architectures being developed in other cancer types, where combining imaging, molecular, and clinical data into a single model consistently improves prediction accuracy over any single input source. Prospective studies embedding VOC testing alongside routine MRI assessment, with expert pathologic diagnosis as the gold standard, would be needed to evaluate this approach.