Knowledge-based interpretable predictive model for survival analysis

#The present embodiments relate to predictive modeling for disease. For example, survival from lung cancer is predicted.

#Survival or survivability from lung cancer, such as non-small cell lung cancer (NSCLC), is relatively low as compared to some other cancers. One common treatment is surgery to resect tumors. Accordingly, various prognosis techniques are directed to patients to be treated with surgery. However, these techniques may not apply to lung cancer patients treated with radiation and/or chemotherapy. An accurate personalized prediction of survival may stratify cancer patients into different risk groups and help in formulating more personalized treatment strategies.

#Patients with stage I-IIIB lung cancer may be treated with curative intent without surgery. Currently, prediction of survival outcome for NSCLC patients treated with chemotherapy and/or radiotherapy is mainly based on clinical factors using TNM staging. However, clinical TNM staging may be inaccurate for survival prediction of non-surgical patients, and alternatives are currently lacking.

#FIG. 1 is a block diagram of one embodiment of a system for knowledge-based interpretable predictive modeling of patients;

#FIG. 2 is a flow chart diagram of one embodiment of a method for knowledge-based interpretable predictive modeling of patients;

#FIG. 3 is an example graphical representation of a predictive model;

#Bayesian, other graphic, or other predictive networks are trained to create personalized predictive models. Unlike some other machine learning based methods such as Support Vector Machines, Bayesian networks naturally reason under uncertainty. The Bayesian network may be used to compute marginal and conditional probability distributions on unobserved nodes and thereby offer a natural way to represent the uncertainties in decision making medical systems. The graphical representation of the network enables a meaningful interpretation of causal or correlated relationships between different attributes and an effective means to reason about new links and graphs. The model may be started based on domain knowledge, such as relationships between variables provided by a physician. This knowledge-based seeding of the model may more effectively create a predictive model.

#The model may be used for predicting side effects and/or survivability with or without treatment for any disease. Below, an example embodiment for predicting two-year survival in lung cancer patients is provided. In this example embodiment, a Bayesian network provides a personalized prediction model for two-year survival in patients with inoperable stage I-III non-small cell lung cancer (NSCLC) treated with radiotherapy. In these patients, generally accepted prognostic and predictive factors include performance status (WHO-PS), weight loss, presence of co-morbidity, radiation dose, tumor size and the use of chemotherapy in addition to radiotherapy. Factors such as gender and age have shown inconsistent results in predicting outcome. 18F-fluorodeoxyglucose positron emission tomography (FDG-PET) and derived factors such as maximal uptake of FDG in the tumor or the number of FDG-PET positive lymph nodes in the mediastinum (PLNS) may be a prognostic variables in NSCLC. However, FDG-PET is not available for all NSCLC patients for a variety of reasons. The predictive models may use the FDG-PET variables when available, but also may be able to predict accurately when these values are missing.

#In data tests using this example, Bayesian networks predict the two-year survival in NSCLC patients as accurately as a Support Vector Machines model, but predict survival more accurately in patients in which one or more of the input features are missing.