Machine Learning Links Functional Status Forecasting to Cost Modeling in PAH
A retrospective cohort study in pulmonary arterial hypertension tested whether machine learning could do more than predict a clinical endpoint. Investigators built a framework that paired future functional class forecasting with downstream health care cost estimation, using baseline data from 181 patients with PAH and 12-month follow-up outcomes.
The work sits at an important intersection in chronic disease care. PAH is uncommon, progressive, and expensive to manage, but the clinical course varies widely from one patient to the next. A model that can anticipate functional decline or improvement may help clinicians think not only about prognosis, but also about how disease trajectory translates into resource use. In this dataset, the central question was whether baseline clinical variables could predict change in World Health Organization functional class and then map that prediction to real-world cost burden.
Predicting WHO Functional Class from Baseline Clinical Data
Six supervised learning methods were trained on baseline variables: LASSO regression, ridge regression, k-nearest neighbors, decision trees, random forest, and gradient boosting. Missing predictor values were handled with median substitution before the dataset was split into training and test sets. Performance was judged with Pearson correlation coefficient, R2, and error metrics including RMSE, MAE, and MedAE under cross-validation and independent test validation.
The strongest performer was k-nearest neighbors, which reached a Pearson correlation of 0.40 and an R2 of 0.16. LASSO regression followed with a correlation of 0.30 and an R2 of 0.09. Error measures were broadly comparable across models, a reminder that no algorithm in this analysis produced a tightly precise forecast. The models captured some signal, but not enough to imply bedside-level certainty for individual patients.
The modest performance is clinically revealing. Functional class in PAH reflects a complex mix of pulmonary vascular resistance, right ventricular adaptation, symptom burden, exercise tolerance, and treatment response. A single baseline snapshot can only approximate that biologic trajectory. Machine learning can absorb nonlinear relationships, but it cannot create information that was never measured in the first place.
One sentence captures the practical limit: the predictive signal was real, but not strong.
Cost Estimates Stayed High Even as Functional Class Improved
Predicted 12-month functional class values were then linked to externally validated, stratified real-world cost estimates. Estimated mean health care costs ranged from US $12,481 to $12,614 per patient per month. Despite a significant improvement in WHO functional class over 12 months, estimated costs remained largely unchanged across strata.
That pattern matters because PAH is often discussed in terms of survival, exercise capacity, and symptom control, while the economic burden can fade into the background. Here, the cost curve did not mirror the functional curve. Patients may improve clinically and still remain expensive to manage, in part because PAH requires ongoing specialty care, surveillance, and long-term therapy. A modest gain in functional class does not necessarily erase the intensity of follow-up or the chronic treatment infrastructure built around the disease.
The economic estimates also suggest that cost pressure persists even when disease status shifts in a favorable direction. In practice, this means a health system cannot assume that better functional class will automatically produce a proportional reduction in spending. Resource planning in PAH may therefore need to account for a relatively fixed high-cost floor, not just changes in short-term clinical status.
Why This Framework Matters for PAH Care Planning
The value of this study lies less in a single best model than in the structure of the framework itself. The investigators showed that predictive modeling can extend beyond purely clinical end points and into system-level forecasting. In a disease where treatment is expensive and clinical heterogeneity is substantial, that combination has obvious appeal for payer planning, care coordination, and service allocation.
Still, the results also define the current boundary of the method. A Pearson correlation of 0.40 is useful for population modeling, but it is not strong enough to support individual patient management on its own. The framework may be more credible as a planning tool than as a direct decision aid. It can help estimate likely shifts in functional class distribution and the economic consequences tied to those shifts, but it does not replace clinical judgment at the point of care.
The study design also leaves important constraints in place. It used retrospective data from a single cohort of 181 patients, and median imputation was applied before the train-test split. While that approach simplifies handling of missing values, it can introduce concerns about information leakage and limit the realism of deployment. External validation in larger, multicenter cohorts will be essential before any operational use in routine care or health system forecasting.
What the Cohort Can and Cannot Support
The most defensible interpretation is straightforward. Baseline clinical data contained enough information to model future functional status in PAH with moderate success, and those predicted trajectories could be paired with cost estimates to approximate ongoing economic burden. The framework is feasible. The performance is only moderate. The cost burden remains stubbornly high.
For clinicians, that means the model is interesting, but not ready to replace established assessment tools. For health systems, it suggests that PAH may benefit from planning models that combine prognosis with economic forecasting rather than treating those questions separately. For researchers, the next step is not more algorithmic complexity alone, but better data density, broader cohorts, and prospective testing of whether integrated prediction actually improves care delivery or spending decisions.
The study leaves one clear unresolved issue: whether a more diverse and larger PAH population would strengthen prediction enough to move this framework from retrospective modeling into practical clinical use.