Machine Learning Tied Functional Class Forecasts to Cost Modeling in Pulmonary Arterial Hypertension
Pulmonary arterial hypertension carries a dual burden: progressive cardiopulmonary limitation and sustained resource use that does not fall neatly in step with symptom improvement. In a retrospective cohort of 181 patients, investigators built an machine learning framework to predict 12-month change in WHO functional class and then translate those predictions into estimated health care costs. The design pushed beyond the usual outcome-prediction exercise. It asked whether a model can forecast not only how patients may function one year later, but also what that future status might cost the health system.
Six supervised algorithms entered the comparison: LASSO regression, ridge regression, k-nearest neighbors, decision trees, random forest, and gradient boosting. Baseline clinical variables served as inputs. Missing values were handled with median substitution before the cohort was split into training and test sets, which keeps the workflow straightforward but also limits how much signal missingness itself could contribute to the model. The target was the relative change in WHO functional class from baseline to 12-month follow-up, a clinically familiar measure in pulmonary hypertension care because it captures worsening or improvement in exertional limitation and daily function.
The best-performing model was k-nearest neighbors.
Its correlation with observed outcomes reached r=0.40 with R2=0.16, a modest result that signals some ability to track functional trajectory while leaving most of the variance unexplained. LASSO regression followed with r=0.30 and R2=0.09. Error metrics were similar across approaches, which suggests no single algorithm broke decisively away from the pack. In practical terms, the cohort produced a framework with signal, but not the kind of precision needed for individual-level decision support in isolation. For a disease as heterogeneous as pulmonary arterial hypertension, where patients may diverge sharply in response to therapy, right heart strain, and comorbidity burden, moderate prediction performance is not surprising.
Functional Improvement Did Not Translate Into Lower Estimated Costs
The economic model created from predicted WHO functional class yielded mean monthly health care costs ranging from US $12,481 to $12,614 per patient. Despite a significant improvement in WHO functional class over 12 months, estimated costs remained largely unchanged across strata. That pattern is clinically important because it separates symptom trajectory from spending trajectory. Patients may move into a better functional category while still requiring ongoing specialty care, advanced therapies, monitoring, and downstream services that keep expenditures elevated.
The result also fits the structure of PAH itself. Functional class reflects the patient’s clinical status, but cost is driven by a broader mix of factors: chronic pharmacologic treatment, repeated follow-up, diagnostic surveillance, and the downstream consequences of a rare disease that typically demands long-term management. A model that predicts class alone may capture only one dimension of the disease burden. Linking that prediction to external cost estimates extends the framework into health system planning, where capacity, reimbursement, and resource allocation depend on more than a single bedside metric.
What the Cohort Can and Cannot Support
The study offers a workable proof of concept, not a ready-to-use clinical tool. A cohort of 181 patients is small by machine learning standards, especially for a disease with meaningful clinical heterogeneity. The retrospective design also limits causal interpretation. Because median imputation occurred before the data split, the workflow may have introduced information leakage risk relative to stricter validation pipelines, a concern that matters when models are meant to generalize beyond the source population. Even so, the overall structure is easy to understand and directly relevant to specialty care: baseline phenotype, predicted future functional class, then projected economic burden.
The choice of endpoints is sensible. WHO functional class remains one of the most recognizable markers in PAH care, and cost estimation is rarely integrated into prediction work at all. Bringing those two outputs into one framework reflects a realistic health services question: which patients are likely to improve, and what does that improvement or lack of improvement mean for spending? In a disease where therapeutic choices are often expensive and longitudinal, even modest predictive tools could support planning if they are validated in larger datasets and embedded in workflows that respect clinical complexity.
Why This Matters for PAH Care Planning
The study’s main contribution lies in its move from pure prediction toward operational forecasting. Many machine learning projects stop at discrimination metrics. Here, the predicted output was converted into a clinically interpretable future functional class and then linked to real-world cost estimates. That sequence matters because hospital systems and payers do not plan around abstract model outputs; they plan around expected visits, treatments, and financial exposure.
Still, the performance ceiling is visible. A correlation of 0.40 leaves substantial room for improvement, and the narrow spread in estimated monthly costs suggests that the cost model may be relatively blunt across functional strata. For clinicians, the message is straightforward: better functional status does not automatically imply a material drop in short-term cost burden. For researchers, the next step is clearer too. Larger multicenter cohorts, external validation, and richer input variables are needed before this kind of framework can support patient-level decision-making or formal health system deployment.
The work shows feasibility, not finality. In pulmonary arterial hypertension, where progression is unpredictable and costs remain stubbornly high, that distinction is enough to justify further study but not enough to change practice on its own.