Merck: cancer care adverse events

From Clinical Data to Actionable Predictions

The study leveraged data from over 600 patients enrolled in multiple clinical trials. These datasets included both baseline characteristics (such as age) and longitudinal measurements (such as lab values tracked over time), reflecting the complexity of real-world clinical care.

To model this, a machine learning framework capable of predicting the likelihood and severity of edema at future clinical visits was proposed. Tree-based models, and specifically Random Forest and Gradient Boosting, were used due to their strong performance on structured clinical data.

A key methodological challenge was incorporating the temporal dimension of patient data. Since standard ML models are not inherently designed for longitudinal inputs, new features capturing recent patient history were engineered, for example by aggregating measurements over different time windows or tracking changes from baseline.

The resulting models achieved high predictive performance, correctly forecasting edema severity in most cases.

Ensuring the Reliability of Predictions

In clinical settings, accuracy alone is not enough: interpretability and reliability are essential.

To ensure that predicted risks correspond to real-world probabilities, probability calibration techniques were applied. This step ensures that a predicted risk truly reflects observed outcomes in similar patients.

Equally important, the study integrated explainability methods using SHAP (Shapley Additive Explanations). These tools allow clinicians and researchers to understand why a model makes a given prediction, both at the population level and for individual patients.

What Drives Edema Risk

Beyond prediction, the approach uncovers clinically meaningful insights. The analysis identified several key factors associated with increased edema risk:

• Serum albumin levels: lower values are strongly linked to higher risk
• Age: older patients are more likely to develop severe edema
• Previous edema status: past events are highly predictive of future ones

These findings are consistent with existing clinical knowledge, reinforcing confidence in the model while also quantifying the relative importance of each factor. The model also supports the understanding of the complex relationships between exposure therapy and adverse events.

Toward Data-Driven Patient Management

This work demonstrates how combining machine learning with explainability can enhance our understanding of treatment safety. By moving beyond static analyses and leveraging longitudinal clinical data, the framework provides a more dynamic view of patient risk over time.

More broadly, it illustrates a shift toward model-informed precision medicine, where data-driven tools support clinicians in anticipating adverse events and tailoring treatments accordingly.

As healthcare data continues to grow in scale and complexity, such approaches will be key to translating information into actionable insights, ultimately improving patient outcomes.

‍

© Image : HockleyM2/peopleimages.com - stock.adobe.com