Who gets fuller prostate-cancer treatment? A small study points to insurance

The medicine has to be within reach

Insurance coverage stood out in a small US study of which patients received recommended additional treatment for advanced prostate cancer. The finding puts access to medicine at the centre of an analysis built with machine learning. But it is a signal from 42 patient records, not a rule that can explain any individual patient’s care.

The peer-reviewed study examined 42 patients treated at Avera Health between 2018 and 2023. Thirty-two received androgen-deprivation therapy plus docetaxel or a newer hormonal therapy within 90 days; 10 did not. Coverage for a newer hormonal therapy was the most consistently selected factor associated with receiving this fuller treatment, appearing in all the model’s outer validation resamples.

That is an important access signal, not proof that insurance caused the treatment decision. Six of the seven patients recorded as lacking coverage for novel hormonal therapy were in the non-intensified group, compared with one of 35 patients recorded as having coverage. With only 10 outcome-negative cases, a handful of records can dominate the pattern. The authors correctly describe the work as exploratory and call for larger, multicentre validation.[1]

The study combined trend data with a separate predictive cohort

The paper has two linked analyses. National treatment patterns came from Pfizer’s Oncology Quality Care Insight Tool, which uses IQVIA anonymised longitudinal patient data. Separately, the prediction work used electronic health records from Avera Health, a US system concentrated in South Dakota and neighbouring states. The institutional cohort contained 42 eligible patients with documented metastatic castration-sensitive disease. Thirty-three lived in South Dakota, 19 were classified rural and 32 were retired.

The outcome was treatment intensification within 90 days of starting androgen-deprivation therapy. Candidate predictors covered age, race and ethnicity, comorbidities, tumour features, performance status, insurance, geography, employment, marital status and smoking. The team excluded direct treatment-receipt variables that could leak the outcome into the model, dropped two heavily missing fields and imputed remaining values with five-nearest-neighbour imputation.[1]

Validation was careful, but the denominator stayed 42

The researchers compared logistic regression, a decision tree, a neural network, random forest, a radial support-vector machine and a stacked ensemble. They used stratified five-fold outer cross-validation repeated 20 times, producing 100 validation resamples. Within each training fold they repeated five-fold inner validation three times for tuning, performed recursive feature elimination and applied synthetic minority oversampling only to training data. This is substantially safer than selecting features and balancing classes before the split.

Resampling, however, does not create new independent patients. Each outer fold held only about eight or nine people, and the minority group contained 10 people in total. Fold-level specificity averaged between 0.42 and 0.53 across models, showing weak discrimination of patients who did not receive intensification. Wide patient-level AUC intervals—such as 0.383 to 0.917 for the random forest figure reported in the abstract—include performance ranging from poor to potentially useful.[1]

The published performance numbers need correction

Readers face a material reporting problem. The abstract says the decision tree had the highest patient-level AUC at 0.675 and assigns the random forest an AUC of 0.650, 85.7% accuracy, 88.2% precision, 93.8% recall and an F1 score of 0.909. The main text repeats those values. Yet Table 3 lists the support-vector machine with the highest AUC at 0.769 and the random forest at 0.722, with 83.3% accuracy and F1 of 0.896. The ROC figure caption presents another set of AUC values.

These may reflect different aggregation steps or an editing mismatch, but the paper does not explain the divergence. A correction should identify which predictions produced each table, curve and abstract result and which values are primary. Until then, it is not responsible to rank the algorithms or quote a single performance figure as definitive. The stable feature-selection result is still descriptive evidence, but even it requires external replication.[1]

The practical finding is a possible barrier, not a patient score

For clinicians and health-system leaders, the actionable question is whether coverage and authorisation delay evidence-based treatment. The national data show newer hormonal therapy use rising from 27% in 2018 to 46% in 2024, while androgen-deprivation monotherapy declined only modestly from 41% to 38%. At Avera, newer hormonal therapy rose from 8% to 88% between 2018 and 2023. Those trends show adoption, but they do not reveal why a particular eligible patient did or did not receive a combination regimen.

A health system does not need a machine-learning score to audit coverage denials, time to authorisation, referral pathways and treatment decisions. It can measure these directly and review exceptions with clinicians and patients. Using the model to predict individual care would risk turning an observed access barrier into an automated expectation. The study did not test a clinical decision-support tool, an intervention to improve access or any effect on survival, quality of life or toxicity.[1]

What would change the assessment

Stronger evidence would include a preregistered, multicentre cohort with enough untreated or delayed cases to estimate specificity and calibration precisely. Researchers should distinguish clinical ineligibility, patient preference, contraindication, affordability, authorisation delay and provider choice rather than compressing them into one 90-day outcome. Models should be locked before testing at new health systems and evaluated across insurance type, race, rurality, age and comorbidity without using post-decision information.

The current paper is most useful as a hypothesis generator and as a reminder that healthcare access can surface as a powerful model feature. It does not establish a causal insurance effect or a deployable predictor. Clarified performance reporting, external validation and direct study of the pathway from prescription to coverage and treatment would be needed before machine learning adds more than a transparent access audit.[1]