New model guides personalized, cost-effective TB treatment

Doctors often must make treatment choices without fast, reliable, or affordable diagnostic tests. In those situations they rely on symptoms and patient history, which can lead to treatments that are either unnecessary or too conservative. Risk prediction models have been developed to estimate the chance that a patient has a particular disease, but those models usually stop at providing a probability and do not take into account what happens next: the health effects, side effects, and costs that flow from giving one treatment rather than another. Mariana R. Neves and colleagues set out to bridge that gap. They developed and evaluated methods that combine risk prediction with decision modeling so that predictions feed directly into decisions that consider both clinical outcomes and economic trade-offs. The goal was to produce personalized, value-based treatment recommendations that work in real clinical settings where diagnostic uncertainty is common, rather than leaving clinicians to choose between a raw probability and a one-size-fits-all rule.

The team created two ways of integrating risk prediction with decision models. In one method, the decision model uses the risk prediction model's binary disease classification to choose a treatment; the alternative method ties predicted risks more directly into the decision process. They tested both approaches on the practical problem of selecting between two treatment regimens for patients with rifampicin-resistant tuberculosis in Moldova. In that application the models explicitly accounted for cost, toxicity, and efficacy associated with each regimen. The results showed that both integration methods produced higher population NMB than the existing standard of care and higher NMB than approaches that apply fixed classification thresholds — for example a 50% cutoff or the threshold that maximizes the Youden's index. The researchers also found that the classification-based approach was less sensitive to whether the model predictions were properly calibrated, a practical advantage when predicted probabilities may be uncertain.

These findings point to a practical way to make prediction tools more useful at the bedside. By folding expected health outcomes and costs into the decision step, clinicians can receive recommendations that reflect the real trade-offs of each option rather than just a risk score. That matters especially in settings with diagnostic uncertainty, where treatment choices carry important consequences for patients and for strained health budgets. Integrating risk prediction with decision modeling offers a principled framework for personalized, value-based treatment decisions: it makes the downstream implications of acting on a prediction explicit, can improve the overall net monetary benefit of care, and supports better use of resources. The comparative robustness of the classification-based approach to calibration issues also suggests a pragmatic path for implementing these methods in real-world clinics where models may not be perfectly tuned.