Category: 3. Business

In an analysis of 7-year follow-up data from the REASSURE study, Bertrand Tombal, MD, emphasized that the findings remain true to the study’s name—reassuring. The results show no significant increase in hematologic toxicity with Radium-223 in patients with metastatic castration-resistant prostate cancer (mCRPC) and bone metastases, while highlighting the critical importance of timing and bone health management.

In an interview with Pharmacy Times, Tombal discussed key takeaways for clinicians and pharmacists, including the need to paid Radium-223 with agents such as denosumab or zoledronic acid and to initiate treatment earlier in the disease course rather than in a palliative setting.

Q: Can you walk us through the key safety findings over this seven-year follow-up and kind of what you found?

Bertrand Tombal, MD: So first I’ll explain why REASSURE was set up. OK, so REASSURE was set up shortly after the ALINA trial because when the data came, there was always that worry that when you use a radioactive compound that goes in the bone marrow of the patient, because that’s where bone metastases are, you could actually increase the rate of hematological malignancies. So that’s why initially REASSURE was done.

So, the first conclusion is that REASSURE is reassuring. We don’t see a significant increase in the toxicity in the hematological toxicity of [Radium-223], so that was very important.

And now it’s close to 1500 patients, so we’ve learned a lot about radium, and basically, we have learned 3 things. The first one is that, yes, the toxicity is hematological, but we rarely have dramatic toxicity like grade 4 or grade 5. Most of the toxicity is limited to grade 1 and grade 3, and it’s really reversible. The second one is linked to the first one to some extent, in that Radium-223 is not a drug that you should use too late as a kind of palliative care setting. If you do that, then you get the toxicity because this patient has bone marrow exhaustion, and you don’t get any therapeutic effect. So that’s something we learned, which is very important. Radium-223 is a drug that has to be given early to metastatic CRPC patients with bone metastasis. If you give it too late, that’s where you’ve got the toxicity, and you don’t have the effect.

And then the third and very important observation is that indeed there is a toxicity, which is not necessarily to the molecule itself but to the fact that you’re tackling bone metastasis. So, if you tackle bone metastasis and you don’t protect your bones, you’re going to get a lot of fractures. There was confusion in the minds of the doctors because drugs like zoledronic acid and denosumab were developed to delay what we call skeletal-related events, which are basically complications of the cancer. When ALINA was published, it showed that it significantly delays [skeletal-related events], so the doctor would say, “No hurry. You don’t need to start this treatment soon.” What happens even with radiotherapy is that we have a lot of frailty and osteoporotic fractures, and you need to protect your patient with this. And so that’s the central message: if you’re using a drug that is heavily targeting the bone, the price to pay for the efficacy is an increased risk of bone fracture, and you absolutely need to combine Radium-223 with a bone-protecting agent.

Is it disruptive? Is it new information? No, it’s been in the guidelines for 15 years. And actually, you could really summarize and say if you give Radium-223, please follow the guidelines. What do the guidelines say? Radium-223 is indicated in patients with metastatic castration-resistant prostate cancer and bone metastasis. The guidelines say zoledronic acid and denosumab should be given to anybody with metastatic CRPC and bone metastasis. So, no surprise, just comply with the guidelines.

What is very interesting is all these messages from REASSURE actually resonate perfectly in the PEACE-3 trial, which is enzalutamide versus enzalutamide plus Radium-223, where we made absolutely the same observation. We haven’t seen any increase in hematological toxicity, and this is a randomized controlled trial.
The toxicity is acceptable. We haven’t seen one single side effect increased by more than 5%, except we see a lot of fractures if we don’t give bone-protecting agents.

So that’s what we learned from this REASSURE trial, the main message being if you believe in Radium-223, it’s not a treatment you give in a palliative care setting. You have to give it quite early. What we see also in REASSURE, not necessarily in the poster represented here, but in other publications, is that the earliest you give it, the higher the chance you’re going to give all 6 cycles. So that, in a nutshell, is the summary of the poster.

Q: From the pharmacist’s perspective, what monitoring do you recommend when a patient is receiving this, especially given the data that you now have?

Tombal: Clearly, 2 things. Any drug that could increase the hematological toxicity—that’s very important. And second, most importantly, because I realize in my hospital [pharmacists] play a critical role in ensuring that the patient is receiving both targeted agents. You know, it’s not listed as a typical drug-drug interaction, so pharmacy software will not pop up because basically if you look at drug-drug interactions, there are no interactions with Radium-223. But the pharmacist is central to making sure that at the time of the prescription of the Radium-223, there is a prescription for 1 bone-targeted agent. We know that in places where the pharmacist has involvement in that quality control monitoring, the rate of incorporation of bone-protecting agents was higher, so that’s very important.

Q: Can you elaborate on how strong the association was between bone-protecting agents and reduced fracture risk, and what recommendations you would make about integrating these agents?

Tombal: Yes. At [the European Association of Urology Congress] last year we published a subset analysis of the first patients who received or did not [bone-protective agents]. And we see that actually, if you take, for instance, [progression-free survival], radiographic progression-free survival, administering a bone-protecting agent was increasing PFS by 15 or 17 months. So, the benefit is huge. And if we look at Radium-223, it was very simple. When we look at the initial part when we didn’t make use of bone protection agents compulsory, in the enzalutamide regimen arm, the rate of fracture was close to 20%. What was interesting is that in the enzalutamide arm alone it was 10%.

And when we administered the bone-protecting agent—in most cases it was denosumab, but zoledronic acid is good as well—we decreased that rate to 3% to 5%. So, my worry when we speak about Radium-223 and bone-protecting agents as a physician is that it’s not only when you give Radium-223; it’s every time your patient has bone metastasis and CRPC. So, my worry is that physicians say, “Oh, he doesn’t need radium, so he doesn’t need a bone-protecting agent.” No, the REASSURE data and the PEACE-3 data show this isn’t true.

Basically, there is a separate message beyond the radium. It’s the importance of bone-protecting agents, whatever drug you receive.

Patients

The study was approved by the Ethics Committee of the First Affiliated Hospital of Soochow University. A total of 880 patients diagnosed with pancreatic cancer between January 1, 2013, and December 31, 2023, were retrospectively collected from two hospitals, including 520 patients from the First Affiliated Hospital of Soochow University and 360 patients from the Second Affiliated Hospital of Soochow University. All patients received follow-up treatment in accordance with the CSCO guidelines for the diagnosis and treatment of pancreatic cancer [16]. The inclusion criteria were: (1) Histopathological diagnosis of pancreatic cancer; (2) Abdominal contrast-enhanced CT examinations performed at our hospital before definitive treatment. The exclusion criteria were: (1) Lack of complete baseline clinical data or follow-up data; (2) Combined with a history of other malignant tumors; (3) Poor quality of contrast-enhanced CT imaging, defined as scans with significant motion artifacts, insufficient contrast enhancement (i.e., poor visualization of pancreatic parenchyma and vessels during the portal venous phase), image blurring affecting tumor boundary visibility, or incomplete coverage of the pancreas within the scanning range. Patients were divided into a training set (n = 616) and a testing set (n = 264) in a 7:3 ratio. The specific selection process for the study cohort is shown in Fig. 1. Demographic characteristics such as gender and age were collected from electronic medical records. Tumor clinical staging was assessed according to the 8th edition of the AJCC (American Joint Committee on Cancer) staging system. Overall survival (OS) was followed up through telephone interviews and inpatient medical records, defined as the time from initial pathological diagnosis to death from any cause or the last known date of survival (Figure S 4). The follow-up was censored on June 1, 2024. OS was binarized as “survival” or “death” for machine learning classification. Death within 1-year, 2-year, and 3-year post-diagnosis was labeled as “death,” while survival at these time points was labeled as “survival.” Patients without follow-up information were already excluded according to the predefined exclusion criteria.

Imaging protocol and preprocessing

All patients underwent contrast-enhanced abdominal CT scans before initiating treatment. Imaging was performed using a 256-slice CT scanner, the GE Revolution CT from GE, USA, at the First Affiliated Hospital of Soochow University, and a 64-slice helical CT scanner, the GE Discovery CT 750 HD, at the Second Affiliated Hospital of Soochow University. Scan parameters were similar at both hospitals, with a tube voltage of 120 kV and a slice thickness and interslice gap of 5 mm. The Second Affiliated Hospital used automatic milliampere modulation, while the First Affiliated Hospital set the tube current between 200 and 450 mAs. Both hospitals used iodixanol 320 as the contrast agent, administered at a dose of approximately 1.4 mL per kilogram of body weight at an injection rate of 3 mL per second. Arterial phase images were acquired using an automatic triggering method, with a threshold of 100 Hounsfield units. Venous phase images were obtained after a 60-second delay.

Portal venous phase contrast-enhanced CT images in DICOM format were imported into 3D Slicer software (version 5.6.1). Guided by a senior radiotherapy oncologist (20 years of experience), a radiotherapy oncologist and a gastroenterologist (each with 3 years of experience) independently delineated tumor boundaries to create regions of interest (ROIs), carefully avoiding adjacent vessels, dilated bile ducts, and pancreatic ducts. Both physicians were blinded to the clinical outcomes prior to the ROI segmentation. If the two physicians produced significantly different ROIs for the same patient, a third radiologist (20 years of experience) reviewed the case and made the final decision in accordance with NCCN (National Comprehensive Cancer Network) guidelines. To assess the reliability and consistency of the ROI delineation, the intraclass correlation coefficient (ICC) was used for validation.

Radiomic feature extraction

Before feature extraction, the images were resampled using 3D Slicer software to a voxel size of 3 × 3 × 3 mm³ to ensure isotropy, thereby minimizing variability due to differences in scanning equipment and protocols. Radiomic features were then extracted using the Radiomics plugin within the 3D Slicer software. A total of 1,037 features were extracted, including 108 original features and 929 filtered features (LoG and wavelet), which were classified into seven categories: shape features, first-order (FO) features, features based on gray-level co-occurrence matrices (GLCM), gray-level dependence matrix (GLDM) features, features based on run-length matrices (RLM), features based on size-zone matrices (SZM), and features based on neighborhood gray-tone difference matrices (NGTDM).

The development and evaluation of models

Radiomics based model

To reduce the dimensionality of features and address the issue of multicollinearity, principal component analysis (PCA) was applied to the extracted radiomic features. We selected principal components that together explained 80% of the cumulative variance to minimize redundancy while retaining the most informative aspects of the original feature set. Subsequently, feature selection was further refined using the Least Absolute Shrinkage and Selection Operator (LASSO) regression, with the optimal λ value determined via 10-fold cross-validation. The final input features were determined by retaining principal components with non-zero coefficients.

The Random Survival Forest (RSF) algorithm, an ensemble tree-based method, is designed to analyze survival time data and predict individual survival probabilities or risks. RSF constructs models by integrating multiple survival trees, providing high stability and robustness while effectively reducing the risk of overfitting. Additionally, it captures complex non-linear relationships and automatically identifies key features closely associated with survival outcomes, thereby enhancing the interpretability of the model [17, 18]. In recent years, RSF has been widely applied to survival analysis problems [19]. In this study, the RSF algorithm was utilized to construct models in the training set for predicting 1-, 2-, and 3-year survival probabilities of patients. The RSF model was implemented using the randomForestSRC package in R. The number of trees (ntree) was set to 345, and the minimum terminal node size (nodesize) was optimized using the tune.nodesize() function, which selects the optimal value based on the out-of-bag (OOB) prediction error. The number of random splits (nsplit) was set to 11. The discriminatory ability of the models was evaluated using the receiver operating characteristic (ROC) curve and its area under the curve (AUC), while prediction consistency was assessed through accuracy measurements.

3D-Densenet based model

Given the importance of capturing three-dimensional structural information in medical imaging, we developed a deep learning model based on a three-dimensional densely connected convolutional network (3D-DenseNet) to predict 1-, 2-, and 3-year survival outcomes in pancreatic cancer patients. The 3D-DenseNet model was implemented using the MONAI (Medical Open Network for AI) framework, which is specifically designed for medical imaging analysis and provides robust deep learning tools optimized for healthcare applications.

The previously segmented tumor regions (ROI) were standardized to a fixed size of 96 × 96 × 96 voxels using linear interpolation to ensure uniform input dimensions and reduce computational complexity. To address the limitations of small sample datasets and enhance model robustness, image preprocessing was performed, including pixel intensity normalization, explicit channel dimension addition, and data type standardization. During the training phase, data augmentation techniques such as random 90-degree rotations, axis flipping, and Gaussian noise injection were applied to improve the model’s generalization capability.

The 3D-DenseNet architecture consisted of four densely connected blocks containing 6, 12, 24, and 16 convolutional layers, respectively. By leveraging dense connectivity, the model reused features from previous layers, improving feature learning efficiency and reducing the number of parameters. To extract both global and local features, a dual-input design (DualInputDenseNet) was adopted, integrating a whole-image branch and an ROI branch. The whole-image branch utilized 3D-DenseNet to extract global features, while the ROI branch employed two 3D convolutional layers, each followed by a ReLU activation function and max pooling operation, to progressively downsample and extract high-level ROI features. The outputs of both branches were concatenated along the channel dimension and passed through fully connected layers for feature fusion and final classification. The model was optimized using the Adam optimizer and trained for 100 epochs with a learning rate of 1 × 10⁻⁴. In each epoch, a batch loader was used to retrieve training data in batches. Cross-entropy loss was employed as the optimization objective, and model performance was monitored using test accuracy computed on the hold-out test set at each epoch. The model checkpoint that achieved the highest test set accuracy was saved for final evaluation. During the testing phase, data augmentation steps were disabled, and only normalization and resizing were applied to ensure stability and consistency in the evaluation process.

After training, the optimal 3D-DenseNet model was used to predict 1-, 2-, and 3-year survival probabilities for patients in both the training and test sets. AUC was calculated to assess the model’s discriminative ability. Additionally, accuracy was evaluated to measure classification consistency.

Fusion model

To further enhance the performance of the prognostic prediction model, this study built a fusion model by performing decision-level fusion of the radiomics model and the 3D-DenseNet model. The predictions from both models were combined with clinical features of patients (gender, age, T stage, N stage, and M stage) to predict the 1-, 2-, and 3-year survival outcomes. This approach aimed to fully leverage the different types of data to enhance the model’s predictive ability, thereby improving its generalizability and accuracy. To identify the optimal fusion model, four different machine learning algorithms were employed, including Logistic Regression, Random Forest, Support Vector Machine (SVM), and Decision Tree. The model performance was evaluated using accuracy and AUC.

After determining the optimal fusion model, patients were stratified based on mortality risk scores predicted by the model. Specifically, the optimal risk cutoff value was identified using X-tile software (version 3.6.1), and all patients were classified into low-risk and high-risk groups accordingly. Kaplan-Meier (KM) survival curves for the high-risk and low-risk groups were then plotted separately in both the training and test cohorts. A log-rank test was used to evaluate the statistical significance of survival differences between the two groups. Figure 2 provide an overview of the overall study workflow.

Statistical analysis

All statistical analyses were performed using SPSS 26.0, R software (version 4.3.2), and Python 2.7.5. The demographic and clinical characteristics between the training and test sets were compared using SPSS 26.0. For continuous variables, an independent sample t-test was applied if the data followed a normal distribution, while the Mann-Whitney U test was used for non-normally distributed variables. Categorical variables were compared using the chi-square test. All statistical tests were two-sided, and a p-value < 0.05 was considered statistically significant. Feature selection, model construction, and evaluation of radiomics-based predictive models were performed using R (version 4.3.2). The “glmnet” package was employed for LASSO regression to select significant radiomic features. The “randomForestSRC” and “survival” packages were used to develop survival prediction models, while the “ggplot2” package was utilized for data visualization. The 3D deep learning model was constructed and evaluated using Python 2.7.5. The remaining general statistical analyses were conducted using R software.

The outage affecting Amazon Web Services, a cloud service provider, points to an overreliance on a centralised system to power the internet.

Amazon’s cloud-computing division has infrastructure set up all around the world, which allows companies to make their products accessible to customers everywhere.

The AWS cloud spans 38 geographic regions, according to its website. But even with this decentralised set-up, there’s an inherent dependence on a small number of key locations. The outages affecting myriad apps and websites this morning were caused by “operational issues” at their datacentre in the US-EAST-1 Region centre in North Virginia, according to AWS. 

“This highlights how reliant we are on centralised services,” says Alp Tolker, founder and director of internet watchdog Netblocks. “The internet isn’t designed to be this way. 

“But what’s emerging is that even within their decentralised infrastructure… one weak link has the ability to take out these different services.”

Toker highlights how these major internet outages often serve as wake-up calls for companies to do things differently. In the case of the AWS outage, there has been a “build fast, fix later” approach. Firms that have been using Amazon’s cloud-computing services to host their data haven’t completed sufficient testing to see what their back-up options are if one datacentre goes down.

“We need to visualise these trees of dependencies, which hasn’t been done formally until now,” says Toker.