Fabien Maldonado

MD, FCCP

Vanderbilt University Medical Center

Professor of Medicine, Thoracic Surgery and Mechanical Engineering

PLCO (Learn more about this study)

PLCO-778

May 7, 2021

Decisions to perform invasive diagnostic procedures follow a probability curve based upon the risk of cancer: insights from lung, ovarian and prostate cancer from the PLCO dataset

Diagnostic risk models are foundational in clinical decision-making regarding lesions suspicious for cancer as they provide an estimate of the probability that a lesion is malignant. The choice to intervene, whether through surgery, minimally invasive biopsy, medication, further imaging, or lifestyle changes, is often a decision made a multidisciplinary team of physicians and caregivers in communication with the patient in a shared decision-making process. The appropriate intervention for most clinical indications is guided by management 'rules' defined by relevant professional societies or government agencies, typically stratifying patients into clinically relevant subgroups. Conventional wisdom suggests that clinical practice follows these guidelines, however data from clinical decisions from lung nodule management in community clinics demonstrates this is not the case(1). The 'all-or-nothing' approach to threshold-based decisions is, in practice, incorrect. In this project, we present a novel approach to threshold-based decision making, the intervention probability curve (IPC). The IPC models the likelihood that a provider will choose the intervention as a continuous function of the risk of the disease. This function can be estimated using professional society guidelines or can be obtained by non-linear regression to historical data on past intervention decisions. The IPC has been fit successfully to a large cohort study, the National Lung Screening Trial (NLST). From this dataset, we extracted the risk of lung cancer as defined by the Mayo Clinic model and the decision of whether to perform an invasive tissue biopsy. The fitted model implied that even at the highest risk, 23% of patients did not undergo biopsy. Modeling clinical decisions using the IPC can have a meaningful impact and can help researchers and clinicians alike. Analysis of the IPC could provide a tool for professional societies or government agencies to assess the congruence of the current guidelines in clinical practice. Information on realistic clinical practice can also help elucidate important questions and inform the best focus for future research efforts. For example, in the NLST dataset, why do the 23% of patients even at highest risk not undergo the invasive diagnostic procedure? This could be due to several reasons: 1) The accuracy of longitudinal imaging is very good, and physicians prefer to wait. 2) The risk of harm by invasive procedure outweighs the benefits of a confirmed diagnosis; among other reasons. The clinical reason for this percentage of high risk not going through invasive procedures is likely a combination of several factors, but each possible reason can help target research efforts. For example, 1 is the case, the research community could focus on improving longitudinal followup time periods, longitudinal models, and encouraging patients to return for followup visits. If 2 is the case, the focus should be on improving the safety and efficacy of the invasive procedures. These are just a few examples, but demonstrate the capability of IPC modeling to assist researchers and clinicians alike in improving disease management strategies.
(1. )Tanner, N.T., et al, 2015. Management of Pulmonary Nodules by Community Pulmonologists. Chest 148, 1405–1414.. doi:10.1378/chest.15-0630

--Compute diagnostic risk scores of candidate patients with suspicious lesions of lung, prostate and ovarian cancer using validated risk models. For example, the project plans to use the ROMA index for ovarian lesions to determine risk of malignancy. Risk scores will be used to fit the Intervention Probability Curve (IPC). The IPC function shape and thresholds can be compared across different risk models for each unique cancer type.

--Fit the Intervention Probability Curve (IPC) model to three specific cancer types: ovarian, prostate and lung to compare probability of intervention across cancer types. Additionally, the probability of intervention will be compared against different levels of intervention to elucidate how realistic clinical action can differ based on risk score and severity of intervention.

-- Analysis of the Intervention Probability Curve (IPC) will provide insight into where improving diagnostic risk models is likely to be most useful. Consider an example where the IPC analysis reveals that physicians are more likely to over-treat a specific condition, and therefore a highly specific biomarker to rule-out patients may be useful.