After the diagnosis of a disease, one major objective is to predict cumulative probabilities of events such as clinical relapse or death from the individual information collected up to a prediction time, usually including biomarker repeated measurements. Several competing estimators have been proposed, mainly from two approaches: joint modelling and landmarking. These approaches differ by the information used, the model assumptions and the complexity of the computational procedures. This paper aims to review the two approaches, precisely define the derived estimators of dynamic predictions and compare their performances notably in case of misspecification. The ultimate goal is to provide key elements for the use of individual dynamic predictions in clinical practice. Prediction of two competing causes of prostate cancer progression from the history of prostate-specific antigen is used as a motivated example. We formally define the quantity to estimate and its estimators, propose techniques to assess the uncertainty around predictions and validate them. We then conduct an in-depth simulation study compare the estimators in terms of prediction error, discriminatory power, efficiency and robustness to model assumptions. We show that prediction tools should be handled with care, in particular by properly specifying models and estimators.

Renal cell carcinoma (RCC) is most often diagnosed at a localized stage, where surgery is the standard of care. Existing prognostic scores provide moderate predictive performance, leading to challenges in establishing follow-up recommendations after surgery and in selecting patients who could benefit from adjuvant therapy. In this study, we developed a model for individual postoperative disease-free survival (DFS) prediction using machine learning (ML) on real-world prospective data. Using the French kidney cancer research network database, UroCCR, we analyzed a cohort of surgically treated RCC patients. Participating sites were randomly assigned to either the training or testing cohort, and several ML models were trained on the training dataset. The predictive performance of the best ML model was then evaluated on the test dataset and compared with the usual risk scores. In total, 3372 patients were included, with a median follow-up of 30 months. The best results in predicting DFS were achieved using Cox PH models that included 24 variables, resulting in an iAUC of 0.81 [IC95% 0.77–0.85]. The ML model surpassed the predictive performance of the most commonly used risk scores while handling incomplete data in predictors. Lastly, patients were stratified into four prognostic groups with good discrimination (iAUC = 0.79 [IC95% 0.74–0.83]). Our study suggests that applying ML to real-world prospective data from patients undergoing surgery for localized or locally advanced RCC can provide accurate individual DFS prediction, outperforming traditional prognostic scores.

Purpose: Triple-negative breast cancer (TNBC) is a biologically and clinically heterogeneous disease, associated with poorer outcomes when compared with other subtypes of breast cancer. Neoadjuvant chemotherapy (NAC) is often given before surgery, and achieving a pathological complete response (pCR) has been associated with patient outcomes. There is thus strong clinical interest in the ability to accurately predict pCR status using baseline data. Materials and Methods: A cohort of 57 TNBC patients who underwent FDG-PET/CT before NAC was analyzed to develop a machine learning (ML) algorithm predictive of pCR. A total of 241 predictors were collected for each patient: 11 clinical features, 11 histopathological features, 13 genomic features, and 206 PET features, including 195 radiomic features. The optimization criterion was the area under the ROC curve (AUC). Event-free survival (EFS) was estimated using the Kaplan–Meier method. Results: The best ML algorithm reached an AUC of 0.82. The features with the highest weight in the algorithm were a mix of PET (including radiomics), histopathological, genomic, and clinical features, highlighting the importance of truly multimodal analysis. Patients with predicted pCR tended to have a longer EFS than patients with predicted non-pCR, even though this difference was not significant, probably due to the small sample size and few events observed (p = 0.09). Conclusions: This study suggests that ML applied to baseline multimodal data can help predict pCR status after NAC for TNBC patients and may identify correlations with long-term outcomes. Patients predicted as non-pCR may benefit from concomitant treatment with immunotherapy or dose intensification.

Long-term blood pressure variability (BPV), an increasingly recognized vascular risk factor, is challenging to analyze. The objective was to assess the impact of BPV modeling on its estimated effect on the risk of stroke. We used data from a secondary stroke prevention trial, PROGRESS (Perindopril Protection Against Stroke Study), which included 6105 subjects. The median number of blood pressure (BP) measurements was 12 per patient and 727 patients experienced a first stroke recurrence over a mean follow-up of 4.3 years. Hazard ratios (HRs) of BPV were estimated from 6 proportional hazards models using different BPV modeling for comparison purposes. The 3 commonly used methods first derived SD of BP measures observed over a given period of follow-up and then used it as a fixed covariate in a Cox model. The 3 more advanced modeling accounted for changes in BP or BPV over time in a single-stage analysis. While the 3 commonly used methods produced contradictory results (for a 5 mmHg increase in BPV, HR=0.75 [95% CI, 0.68–0.82], HR=0.99 [0.91–1.08], HR=1.19 [1.10–1.30]), the 3 more advanced modeling resulted in a similar moderate positive association (HR=1.08 [95% CI, 0.99–1.17]), whether adjusted for BP at randomization or mean BP over the follow-up. The method used to assess BPV strongly affects its estimated effect on the risk of stroke, and should be chosen with caution. Further methodological developments are needed to account for the dynamics of both BP and BPV over time, to clarify the specific role of BPV.

Background: Differentiating benign from malignant renal tumors is important for patient management, and it may be improved by quantitative CT features analysis including radiomic. Purpose: This study aimed to compare performances of machine learning models using bio-clinical, conventional radiologic and 3D-radiomic features for the differentiation of benign and malignant solid renal tumors using pre-operative multiphasic contrast-enhanced CT examinations. Materials and methods: A unicentric retrospective analysis of prospectively acquired data from a national kidney cancer database was conducted between January 2016 and December 2020. Histologic findings were obtained by robotic-assisted partial nephrectomy. Lesion images were semi-automatically segmented, allowing for a 3D-radiomic features extraction in the nephrographic phase. Conventional radiologic parameters such as shape, content and enhancement were combined in the analysis. Biological and clinical features were obtained from the national database. Eight machine learning (ML) models were trained and validated using a ten-fold cross-validation. Predictive performances were evaluated comparing sensitivity, specificity, accuracy and AUC. Results: A total of 122 patients with 132 renal lesions, including 111 renal cell carcinomas (RCCs) (111/132, 84%) and 21 benign tumors (21/132, 16%), were evaluated (58 +/− 14 years, men 74%). Unilaterality (100/111, 90% vs. 13/21, 62%; p = 0.02), necrosis (81/111, 73% vs. 8/21, 38%; p = 0.02), lower values of tumor/cortex ratio at portal time (0.61 vs. 0.74, p = 0.01) and higher variation of tumor/cortex ratio between arterial and portal times (0.22 vs. 0.05, p = 0.008) were associated with malignancy. A total of 35 radiomics features were selected, and “intensity mean value” was associated with RCCs in multivariate analysis (OR = 0.99). After ten-fold cross-validation, a C5.0Tree model was retained for its predictive performances, yielding a sensitivity of 95%, specificity of 42%, accuracy of 87% and AUC of 0.74. Conclusion: Our machine learning-based model combining clinical, radiologic and radiomics features from multiphasic contrast-enhanced CT scans may help differentiate benign from malignant solid renal tumors.