The aim of the study was to validate a predictive biomarker machine learning model for the classification of Parkinson's disease (PD) and age-matched controls (AMC), based on bioelement abundance in the cerebrospinal fluid (CSF). For this multicentric trial, participants were enrolled from four different centers. CSF was collected according to standardized protocols. For bioelement determination, CSF samples were subjected to inductively coupled plasma mass spectrometry. A predefined Support Vector Machine (SVM) model, trained on a previous discovery cohort was applied for differentiation, based on the levels of six different bioelements. 82 PD patients, 68 age-matched controls and 7 additional Normal Pressure Hydrocephalus (NPH) patients were included to validate a predefined SVM model. Six differentiating elements (As, Fe, Mg, Ni, Se, Sr) were quantified. Based on their levels, SVM was successfully applied to a new local cohort (AUROC 0.76, Sensitivity 0.80, Specificity 0.83), without taking any additional features into account. The same model did not discriminate PD and AMCs / NPH from three external cohorts, likely due to center effects. However, discrimination was possible in cohorts with a full elemental data set, now using center-specific discovery cohorts and a cross validated approach (AUROC 0.78 and 0.88, respectively). Pooled PD CSF iron levels showed a clear correlation with disease duration (p = .0001). In summary, bioelemental CSF patterns, obtained by mass spectrometry and integrated into a predictive model yield the potential to facilitate the differentiation of PD and AMC. Center-specific biases interfere with application in external cohorts. This must be carefully addressed using center-defined, local reference values and models.

Federated learning is a renowned technique for utilizing decentralized data while preserving privacy. However, real-world applications often face challenges like partially labeled datasets, where only a few locations have certain expert annotations, leaving large portions of unlabeled data unused. Leveraging these could enhance transformer architectures’ ability in regimes with small and diversely annotated sets. We conduct the largest federated cardiac CT analysis to date (n = 8, 104) in a real-world setting across eight hospitals. Our two-step semi-supervised strategy distills knowledge from task-specific CNNs into a transformer. First, CNNs predict on unlabeled data per label type and then the transformer learns from these predictions with label-specific heads. This improves predictive accuracy and enables simultaneous learning of all partial labels across the federation, and outperforms UNet-based models in generalizability on downstream tasks. Code and model weights are made openly available for leveraging future cardiac CT analysis.

This study evaluates the diagnostic performance of radiomic features and machine learning algorithms for renal tumor subtype assessment in venous computed tomography (CT) studies from clinical routine. Patients undergoing surgical resection and histopathological assessment of renal tumors at a tertiary referral center between 2012 and 2019 were included. Preoperative venous-phase CTs from multiple referring imaging centers were segmented, and standardized radiomic features extracted. After preprocessing, class imbalance handling, and feature selection, machine learning algorithms were used to predict renal tumor subtypes using 10-fold cross validation, assessed as multiclass area under the curve (AUC). In total, n = 201 patients were included (73.7% male; mean age 66 ± 11 years), with n = 131 clear cell renal cell carcinomas (ccRCC), n = 29 papillary RCC, n = 11 chromophobe RCC, n = 16 oncocytomas, and n = 14 angiomyolipomas (AML). An extreme gradient boosting algorithm demonstrated the highest accuracy (multiclass area under the curve (AUC) = 0.72). The worst discrimination was evident for oncocytomas vs. AML and oncocytomas vs. chromophobe RCC (AUC = 0.55 and AUC = 0.45, respectively). In sensitivity analyses excluding oncocytomas, a random forest algorithm showed the highest accuracy, with multiclass AUC = 0.78. Radiomic feature analyses from venous-phase CT acquired in clinical practice with subsequent machine learning can discriminate renal tumor subtypes with moderate accuracy. The classification of oncocytomas seems to be the most complex with the lowest accuracy.

Aims: Identification of high-risk patients and individualized decision support based on objective criteria for rapid discharge after transcatheter aortic valve implantation (TAVI) are key requirements in the context of contemporary TAVI treatment. This study aimed to predict 30-day mortality following TAVI based on machine learning (ML) using data from the German Aortic Valve Registry. Methods and results: Mortality risk was determined using a random forest ML model that was condensed in the newly developed TAVI Risk Machine (TRIM) scores, designed to represent clinically meaningful risk modelling before (TRIMpre) and in particular after (TRIMpost) TAVI. Algorithm was trained and cross-validated on data of 22 283 patients (729 died within 30 days post-TAVI) and generalisation was examined on data of 5864 patients (146 died). TRIMpost demonstrated significantly better performance than traditional scores [C-statistics value, 0.79; 95% confidence interval (CI)] [0.74; 0.83] compared to Society of Thoracic Surgeons (STS) with C-statistics value 0.69; 95%-CI [0.65; 0.74]). An abridged (aTRIMpost) score comprising 25 features (calculated using a web interface) exhibited significantly higher performance than traditional scores (C-statistics value, 0.74; 95%-CI [0.70; 0.78]). Validation on external data of 6693 patients (205 died within 30 days post-TAVI) of the Swiss TAVI Registry confirmed significantly better performance for the TRIMpost (C-statistics value 0.75, 95%-CI [0.72; 0.79]) compared to STS (C-statistics value 0.67, CI [0.63; 0.70]). Conclusion: TRIM scores demonstrate good performance for risk estimation before and after TAVI. Together with clinical judgement, they may support standardised and objective decision-making before and after TAVI.

High-throughput technologies are becoming increasingly important in discovering prognostic biomarkers and in identifying novel drug targets. With Mammaprint, Oncotype DX, and many other prognostic molecular signatures breast cancer is one of the paradigmatic examples of the utility of high-throughput data to deliver prognostic biomarkers, that can be represented in a form of a rather short gene list. Such gene lists can be obtained as a set of features (genes) that are important for the decisions of a Machine Learning (ML) method applied to high-dimensional gene expression data. Several studies have identified predictive gene lists for patient prognosis in breast cancer, but these lists are unstable and have only a few genes in common. Instability of feature selection impedes biological interpretability: genes that are relevant for cancer pathology should be members of any predictive gene list obtained for the same clinical type of patients. Stability and interpretability of selected features can be improved by including information on molecular networks in ML methods. Graph Convolutional Neural Network (GCNN) is a contemporary deep learning approach applicable to gene expression data structured by a prior knowledge molecular network. Layer-wise Relevance Propagation (LRP) and SHapley Additive exPlanations (SHAP) are methods to explain individual decisions of deep learning models. We used both GCNN+LRP and GCNN+SHAP techniques to construct feature sets by aggregating individual explanations. We suggest a methodology to systematically and quantitatively analyze the stability, the impact on the classification performance, and the interpretability of the selected feature sets. We used this methodology to compare GCNN+LRP to GCNN+SHAP and to more classical ML-based feature selection approaches. Utilizing a large breast cancer gene expression dataset we show that, while feature selection with SHAP is useful in applications where selected features have to be impactful for classification performance, among all studied methods GCNN+LRP delivers the most stable (reproducible) and interpretable gene lists.