Predictive Models Could Address Missing Data in Advanced Prostate Cancer

Eastern Cooperative Oncology Group (ECOG) performance status plays an important role in treatment selection, prognosis, and outcomes assessment in advanced prostate cancer. Higher ECOG scores have been associated with an increased mortality risk among patients with metastatic castration-resistant prostate cancer.¹ Despite its clinical value, performance status is frequently unavailable at the specific point in a patient’s treatment journey required for real-world evidence studies.

Investigators at Ontada evaluated whether routinely collected electronic health record (EHR) data could serve as proxies for missing baseline ECOG performance status. Their retrospective study included adults with advanced prostate cancer treated within The US Oncology Network and compared a machine-learning approach with a conventional multiple-imputation model.²

ECOG Documentation Gaps Extend Beyond Missing Data

ECOG performance status is based on a clinician’s assessment of a patient’s ability to complete daily activities; however, the score may be documented in an unstructured clinical note, entered elsewhere in the EHR, or not updated near the initiation of a treatment being evaluated.

“Although it is central to decision-making, it’s not consistently documented in the EHR,” Carole R. Berini, PhD, research scientist at Ontada, explained in an interview with Pharmacy Times. “Sometimes it’s in the notes, sometimes it might be somewhere else. It’s not necessarily updated around the initiation of treatments that we want to study.”

The timing of documentation is particularly important in observational research. According to Saamir Pasha, MPH, senior biostatistician at Ontada, patients may have ECOG information elsewhere in their records even when a baseline score is unavailable for a study.

“We seem to have very good coverage for any ECOG status,” Pasha said. “However, for our studies, we’re usually focused on a very particular point in time in that patient’s history. ECOG isn’t recorded routinely or doesn’t align with the cadence that we would like for our particular studies.”

These gaps can affect the validity of downstream analyses because ECOG performance status is associated with prognosis. Researchers who exclude patients without a documented baseline score may reduce both sample size and statistical power. More importantly, documentation may not be missing randomly.

“If the analysis only includes patients with a documented ECOG, it not only reduces the sample size, but it could also potentially introduce some selection bias,” Berini explained. “That might result in evidence that is less representative of all the patients who are routinely treated in community oncology practice.”

Previous studies have similarly described the difficulty of obtaining ECOG status from structured EHR data and have investigated natural-language processing, imputation, and proxy measures to recover this information.³˒⁴

Investigators Compare 2 Modeling Approaches

The investigators required patients to have documented ECOG status near treatment initiation during model development, allowing the recorded scores to serve as the outcome against which predictions were evaluated. Candidate predictors were selected from structured EHR fields using published literature and clinical expertise.²

The study compared automated machine learning, or AutoML, with multiple imputation using backward-elimination logistic regression. The AutoML workflow tested several candidate algorithms and preprocessing decisions instead of requiring investigators to select a single model in advance.

“In plain terms, this AutoML workflow helps us automate the repetitive parts of model development,” Pasha said. These steps included dividing the data into training and validation sets, addressing missing values, encoding variables, evaluating model families, and testing hyperparameters.

A gradient-boosting machine was identified as the leading machine-learning model. Pasha explained that this approach was well suited to the clinical question because it could capture nonlinear relationships and interactions among factors such as anemia, low albumin, and renal disease.

The machine-learning model achieved an area under the receiver operating characteristic curve (AUROC) of about 0.77 and an F1 score of 0.37 for poorer ECOG status. The multiple-imputation model produced a similar AUROC of 0.74 in the training data.²

Several predictors were consistent with factors clinicians might associate with functional vulnerability. Older age, lower body mass index, lower hemoglobin, more advanced disease, and renal disease were among the signals associated with poorer ECOG status.

“We didn’t just use AutoML as a black-box answer and stop there,” Pasha explained. “We used it to help us search the modeling space efficiently while conducting a parallel statistical model and clinical review to help us pressure-test these signals.”

This parallel approach allowed the researchers to determine whether the more flexible machine-learning model and the conventional statistical model identified similar relationships. “The goal is not just to predict, but to have a defensible, clinically interpretable estimate.”

High Overall Accuracy Conceals Low Sensitivity

In a validation cohort of nearly 1200 patients, the multiple-imputation model achieved an accuracy of 89.4% and specificity of 99.1%; however, sensitivity for identifying patients with poorer ECOG status was only 5.1%. Approximately 10% of the population had documented poorer ECOG status, creating a class-imbalance problem.² This imbalance means the model could achieve high overall accuracy by correctly identifying patients in the much larger group with better performance status while failing to detect many patients in the smaller poor-status group.

“This is the real important limitation,” Pasha said. “Only about 10% of our population had a documented poor ECOG status, so it’s very reasonable to expect a model to miss these outcomes more often than not.”

The findings reinforce the need to examine sensitivity, specificity, calibration, and subgroup performance instead of relying on overall accuracy alone. A model intended to expand representation in research would provide limited value if it continued to overlook the patients with the greatest functional vulnerability.

Berini emphasized that validation must evaluate more than agreement between predicted and recorded ECOG status. “The standard really becomes: Does the predicted ECOG preserve clinical and outcome relationships for the patients? Once we can validate that, then it becomes usable in the broader sense.”

More Complete Data Could Improve Representation

Patients with documented ECOG scores may differ systematically from those without scores recorded during a study’s baseline window. Conditioning study inclusion on documentation could therefore alter the population represented in an analysis. “The concern here with the bias is not simply that we lose patients; it is that we lose a certain type of patient,” Berini said.

A validated proxy could allow researchers to retain patients who would otherwise be excluded, creating real-world evidence that better reflects community oncology practice. However, researchers would still need to determine which populations are most affected by missing ECOG data and whether the proxy successfully restores representation for those groups.

Prior research has developed ECOG proxy models using EHR or claims data in advanced non–small cell lung cancer and across multiple tumor types.^3,4 The current study extends this work into advanced prostate cancer and a community oncology population.²

Implications for Oncology Pharmacists

Performance status can influence treatment intensity, supportive care requirements, toxicity monitoring, and medication management. More complete functional-status data could help oncology pharmacists judge whether evidence applies to the patients they encounter in practice.

“If the evidence is more comprehensive and representative, it becomes more interpretable in terms of how the treatment performs across a range of different patients,” Berini said. “A better characterization could help pharmacists assess whether the evidence is fit for the patients they actually see in practice.”

Before these proxies could support comparative-effectiveness or regulatory research, they would require transparent derivation and external validation across health systems, patient populations, tumor types, and intended applications. Sensitivity analyses and clinical review of cases in which predicted and documented ECOG scores disagree would also be important. FDA guidance emphasizes that real-world data must be relevant and reliable for their proposed regulatory purpose.⁵

Improving routine ECOG documentation remains an important upstream goal. Pasha suggested that greater awareness and more seamless EHR workflows could increase the availability of performance-status data without adding substantial burden in busy oncology practices.

For now, the results establish feasibility rather than readiness for routine research use. The models detected clinically coherent patterns, but their limited ability to identify patients with poorer ECOG status must be addressed before proxy-derived scores can reliably broaden the populations represented in real-world oncology studies.

REFERENCES

1. Assayag J, Kim C, Chu H, Webster J. The prognostic value of Eastern Cooperative Oncology Group performance status on overall survival among patients with metastatic prostate cancer: a systematic review and meta-analysis. Front Oncol. 2023;13:1194718. Published 2023 Dec 15. doi:10.3389/fonc.2023.1194718

2. Abstracts of ISPEs 2026, 42nd International Conference, August 29–September 2, 2026, Milan. Pharmacoepidemiol Drug Saf, 2026;35(Suppl 3):e70409. doi:10.1002/pds.70409

3. Sadetsky N, Chuo CY, Davidoff AJ. Development and evaluation of a proxy for baseline ECOG PS in advanced non-small cell lung cancer, bladder cancer, and melanoma: An electronic health record study. Pharmacoepidemiol Drug Saf. 2021;30(9):1233-1241. doi:10.1002/pds.5309

4. Sheffield KM, Bowman L, Smith DM, et al. Development and validation of a claims-based approach to proxy ECOG performance status across ten tumor groups. J Comp Eff Res. 2018;7(3):193-208. doi:10.2217/cer-2017-0040

5. US Food and Drug Administration. Real-world evidence. Updated June 3rd, 2026. Accessed September 14, 2026. https://www.fda.gov/science-research/science-and-research-special-topics/real-world-evidence