Nicolas Captier’s research while affiliated with Mines Paris, PSL University and other places

Immunotherapy is improving the survival of patients with metastatic non-small cell lung cancer (NSCLC), yet reliable biomarkers are needed to identify responders prospectively and optimize patient care. In this study, we explore the benefits of multimodal approaches to predict immunotherapy outcome using multiple machine learning algorithms and integration strategies. We analyze baseline multimodal data from a cohort of 317 metastatic NSCLC patients treated with first-line immunotherapy, including positron emission tomography images, digitized pathological slides, bulk transcriptomic profiles, and clinical information. Testing multiple integration strategies, most of them yield multimodal models surpassing both the best unimodal models and established univariate biomarkers, such as PD-L1 expression. Additionally, several multimodal combinations demonstrate improved patient risk stratification compared to models built with routine clinical features only. Our study thus provides evidence of the superiority of multimodal over unimodal approaches, advocating for the collection of large multimodal NSCLC datasets to develop and validate robust and powerful immunotherapy biomarkers.

In the analysis of medical data with censored outcomes, identifying the optimal machine learning pipeline is a challenging task, often requiring extensive preprocessing, feature selection, model testing, and tuning. To investigate the impact of the choice of pipeline on prediction performance, we evaluated 9 machine learning models on 71 medical datasets with censored targets. Only the decision tree model was consistently underperforming, while the other 8 models performed similarly across datasets, with little to no improvement from preprocessing optimization and hyperparameter tuning. Interestingly, more complex models did not outperform simpler ones, and reciprocally. ICARE, a straightforward model univariately learning only the sign of each feature instead of a weight, demonstrated similar performance to other models across most datasets while exhibiting lower overfitting, particularly in high-dimensional datasets. These findings suggest that using the ICARE model to build signatures between centers could improve reproducibility. Our findings also challenge the traditional approach of extensive model testing and tuning to improve performance.

The survival of patients with metastatic non-small cell lung cancer (NSCLC) has been increasing with immunotherapy, yet efficient biomarkers are still needed to optimize patient care. In this study, we explored the benefits of multimodal approaches to predict immunotherapy outcome using multiple machine learning algorithms and integration strategies. We leveraged a novel multimodal cohort of 317 metastatic NSCLC patients treated with first-line immunotherapy, collecting at baseline positron emission tomography images, digitized pathological slides, bulk transcriptomic profiles, and clinical information. Most integration strategies investigated yielded multimodal models surpassing both the best unimodal models and established univariate biomarkers, such as PD-L1 expression. Additionally, several multimodal combinations demonstrated improved patient risk stratification compared to models built with routine clinical features only. Our study thus provided new evidence of the superiority of multimodal over unimodal approaches, advocating for the collection of large multimodal NSCLC cohorts to develop and validate robust and powerful immunotherapy biomarkers.

Explaining the decisions made by a radiomic model is of significant interest, as it can provide valuable insights into the information learned by complex models and foster trust in well-performing ones, thereby facilitating their clinical adoption. Promising radiomic approaches that aggregate information from multiple regions within an image currently lack suitable explanation tools that could identify the regions that most significantly influence their decisions. Here we present a model- and modality-agnostic tool (RadShap, https://github.com/ncaptier/radshap), based on Shapley values, that explains the predictions of multiregion radiomic models by highlighting the contribution of each individual region. Methods: The explanation tool leverages Shapley values to distribute the aggregative radiomic model's output among all the regions of interest of an image, highlighting their individual contribution. RadShap was validated using a retrospective cohort of 130 patients with advanced non-small cell lung cancer undergoing first-line immunotherapy. Their baseline PET scans were used to build 1,000 synthetic tasks to evaluate the degree of alignment between the tool's explanations and our data generation process. RadShap's potential was then illustrated through 2 real case studies by aggregating information from all segmented tumors: the prediction of the progression-free survival of the non-small cell lung cancer patients and the classification of the histologic tumor subtype. Results: RadShap demonstrated strong alignment with the ground truth, with a median frequency of 94% for consistently explained predictions in the synthetic tasks. In both real-case studies, the aggregative models yielded superior performance to the single-lesion models (average [±SD] time-dependent area under the receiver operating characteristic curve was 0.66 ± 0.02 for the aggregative survival model vs. 0.55 ± 0.04 for the primary tumor survival model). The tool's explanations provided relevant insights into the behavior of the aggregative models, highlighting that for the classification of the histologic subtype, the aggregative model used information beyond the biopsy site to correctly classify patients who were initially misclassified by a model focusing only on the biopsied tumor. Conclusion: RadShap aligned with ground truth explanations and provided valuable insights into radiomic models' behaviors. It is implemented as a user-friendly Python package with documentation and tutorials, facilitating its smooth integration into radiomic pipelines.

The normalized distances from the hot spot of radiotracer uptake (SUVmax) to the tumor centroid (NHOC) and to the tumor perimeter (NHOP) have recently been suggested as novel PET features reflecting tumor aggressiveness. These biomarkers characterizing the shift of SUVmax toward the lesion edge during tumor progression have been shown to be prognostic factors in breast and non-small cell lung cancer (NSCLC) patients. We assessed the impact of imaging parameters on NHOC and NHOP, their complementarity to conventional PET features, and their prognostic value for advanced-NSCLC patients. Methods: This retrospective study investigated baseline [18F]FDG PET scans: cohort 1 included 99 NSCLC patients with no treatment-related inclusion criteria (robustness study); cohort 2 included 244 NSCLC patients (survival analysis) treated with targeted therapy (93), immunotherapy (63), or immunochemotherapy (88). Although 98% of patients had metastases, radiomic features including SUVs were extracted from the primary tumor only. NHOCs and NHOPs were computed using 2 approaches: the normalized distance from the localization of SUVmax or SUVpeak to the tumor centroid or perimeter. Bland-Altman analyses were performed to investigate the impact of both spatial resolution (comparing PET images with and without gaussian postfiltering) and image sampling (comparing 2 voxel sizes) on feature values. The correlation of NHOCs and NHOPs with other features was studied using Spearman correlation coefficients (r). The ability of NHOCs and NHOPs to predict overall survival (OS) was estimated using the Kaplan-Meier method. Results: In cohort 1, NHOC and NHOP features were more robust to image filtering and to resampling than were SUVs. The correlations were weak between NHOCs and NHOPs (r ≤ 0.45) and between NHOCs or NHOPs and any other radiomic features (r ≤ 0.60). In cohort 2, the patients with short OS demonstrated higher NHOCs and lower NHOPs than those with long OS. NHOCs significantly distinguished 2 survival profiles in patients treated with immunotherapy (log-rank test, P < 0.01), whereas NHOPs stratified patients regarding OS in the targeted therapy (P = 0.02) and immunotherapy (P < 0.01) subcohorts. Conclusion: Our findings suggest that even in advanced NSCLC patients, NHOC and NHOP features pertaining to the primary tumor have prognostic potential. Moreover, these features appeared to be robust with respect to imaging protocol parameters and complementary to other radiomic features and are now available in LIFEx software to be independently tested by others.

e21164 Background: Overall survival of patients with metastatic non-small cell lung cancer (NSCLC) has increased with the use of anti-PD-1 immune checkpoint inhibitors. However, the duration of response remains highly variable between patients, and only 20-30% of patients are alive at 2 years. Thus, new biomarkers for predicting response to treatment and patient outcomes are still needed to guide therapeutic decision. In this study, we retrospectively investigated multimodal approaches that might improve the limited predictive power of clinical data. Methods: We studied a cohort of 317 patients with metastatic NSCLC treated with first-line immune checkpoint inhibitors alone or combined with platinum-based chemotherapy. Clinical data were collected for all patients, pathological slides (HES and PD-L1 staining) and baseline 18-FDG PET/CT scans were available in 237 and 130 patients respectively. An automatic cell type detection algorithm was applied to each pathological slide and pathomic features were extracted from the resulting annotations. After semi-automated segmentation of all tumor foci in the PET/CT scans, radiomic features were calculated for each tumor lesion and aggregated across all the lesions of each patient. Prognostic models were built using random forest and XGboost classifiers to predict patient survival at 12 months based on 1) features from single modalities (clinical, pathomic, or radiomic), 2) features from multiple modalities, where early fusion and late fusion strategies were investigated. The models were trained and tested with cross-validation and their performances were established using the area under the ROC curve (AUC) computed on the same 88 test patients for whom all the modalities were available. Results: Unimodal strategies yielded AUC of 0.62 ± 0.08 (1 std), 0.64 ± 0.07, 0.59 ± 0.08 for clinical, radiomic and pathomic features respectively. With late fusion, bimodal models consistently outperformed the clinical model, with the combination of radiomic and clinical features giving the best performance (AUC = 0.67 ± 0.07). The trimodal model outperformed all other modality combinations with an AUC of 0.69 ± 0.07; in particular, it was significantly superior to the clinical model (p-value < 0.001, paired t-test). The early fusion experiments confirmed the superiority of every bimodal approach over the clinical model. However, the trimodal model did not outperform the best bimodal model with early fusion. Validation will be performed on independent cohorts from external centers. Conclusions: Our study highlighted the potential of multimodal approaches for predicting the outcome of metastatic NSCLC patients treated with immunotherapy. Models integrating medical images and pathological slides usually collected from routine care outperformed a model trained on clinical data alone.

We developed BIODICA, an integrated computational environment for application of Independent Component Analysis (ICA) to bulk and single-cell molecular profiles, interpretation of the results in terms of biological functions and correlation with metadata. The computational core is the novel Python package stabilized-ica which provides interface to several ICA algorithms, a stabilization procedure, meta-analysis and component interpretation tools. BIODICA is equipped with a user-friendly graphical user interface, allowing non-experienced users to perform the ICA-based omics data analysis. The results are provided in interactive ways, thus facilitating communication with biology experts. Availability and Implementation BIODICA is implemented in Java, Python and JavaScript. The source code is freely available on GitHub under the MIT and the GNU LGPL licenses. BIODICA is supported on all major operating systems. Url https://sysbio-curie.github.io/biodica-environment/