Knowledge-Based Interpretable Predictive Model for Survival Analysis

• US 20100057651A1
• Filed: 07/21/2009
• Published: 03/04/2010
• Est. Priority Date: 09/03/2008
• Status: Active Grant

• Current Assignee: Siemens Medical Solutions USA Incorporated (Siemens AG)
• Original Assignee: Maastro Clinic, Siemens Medical Solutions USA Incorporated (Siemens AG)
• Inventors: Dekker, Andreas Lubbertus Aloysius Johannes, Komati, Kartik Jayasurya, Lambin, Philippe, Yu, Shipeng, Fung, Glenn, Dehing-Oberije, Cary
• Granted Patent: US 8,078,554 B2
• Time in Patent Office: Days
• US Class Current: 706/12
• CPC Class Codes: A61N 2005/1041 using a library of previously administered radiation treatment applied to other patients
• CPC Class Codes: G06N 7/01 Probabilistic graphical models, e.g. probabilistic networks

Abstract

Knowledge-based interpretable predictive modeling is provided. Expert knowledge is used to seed training of a model by a machine. The expert knowledge may be incorporated as diagram information, which relates known causal relationships between predictive variables. A predictive model is trained. In one embodiment, the model operates even with a missing value for one or more variables by using the relationship between variables. For application, the model outputs a prediction, such as the likelihood of survival for two years of a lung cancer patient. A graphical representation of the model is also output. The graphical representation shows the variables and relationships between variables used to determine the prediction. The graphical representation is interpretable by a physician or other to assist in understanding.

Claims

1. A system for knowledge-based interpretable predictive modeling of patients, the system comprising:

an input configured to receive patient information representing a characteristic of a first patient;

a processor configured to apply a graphical model as a function of the patient information, the model configured to output a prediction for the first patient; and

a display configured to output an image, the image comprising a graphical representation of the graphical model and the prediction.

2. The system of claim 1 wherein the patient is a lung cancer patient, and wherein the characteristic comprises at least two of tumor load, T-stage, N-stage, number of lymph node stations, and who performance.

3. The system of claim 2 wherein the graphical model comprises a Bayesian network with links from a tumor load node to a T-stage node, a N-stage, and a survival node, from the T-stage node to a number of lymph node stations node, from the N-stage node to the number of lymph node stations node, from a who score node to the survival node, and from the number of lymph node stations node to the survival node.

4. The system of claim 3 wherein the graphical representation shows the nodes and the links.

5. The system of claim 1 wherein the graphical model comprises a model for non-small cell lung cancer.

6. The system of claim 1 wherein the graphical model is a machine-learned Bayesian network with a seed for training the graphical model provided from an expert in a medical domain and the seed for training representing relationships between variables.

7. The system of claim 1 wherein the output of the model represents a probability for survival through a period.

8. The system of claim 1 wherein the patient information represents additional characteristics, and wherein the graphical representation includes nodes associated with respective characteristics, a value for each characteristic included with the respective node, and includes links between some nodes and not between other nodes, the links representing relationships of the graphical model.

9. In a computer readable storage medium having stored therein data representing instructions executable by a programmed processor for knowledge-based interpretable predictive modeling of patients, the instructions comprising:

receiving diagram information representing relationships between variables of a disease;

seeding a predictive model with the diagram information;

training the predictive model, as seeded with the diagram information, with training data, the data comprising values for the variables of the disease; and

generating a graphical representation of the predictive model after the training, the graphical representation showing at least one of the relationships.

10. The computer readable storage medium of claim 9 further comprising instructions for:

applying the predictive model after the training to patient information for a patient; and

displaying a prediction output by the predictive model as a result of the applying and displaying the graphical representation.

11. The computer readable storage medium of claim 9 wherein the disease is lung cancer, and wherein the variables comprise at least two of tumor load, T-stage, N-stage, number of lymph node stations, who performance, and survival, the predictive model trained to predict the survival.

12. The computer readable storage medium of claim 11 wherein the predictive model comprises a Bayesian network, and wherein the diagram information includes links from a tumor load node to a T-stage node, a N-stage, and a survival node, from the T-stage node to a number of lymph node stations node, from the N-stage node to the number of lymph node stations node, from a who score node to the survival node, and from the number of lymph node stations node to the survival node.

13. The computer readable storage medium of claim 12 wherein the graphical representation shows the nodes and the relationships as links between the nodes.

14. The computer readable storage medium of claim 9 wherein the diagram information is received from an expert in a medical domain.

15. The computer readable storage medium of claim 9 wherein the graphical representation includes nodes associated with respective variables and includes links between some nodes and not between other nodes.

16. A method for knowledge-based interpretable predictive modeling of patients, the method comprising:

training, with machine training using training data for a plurality of previous cancer patients, a graphic model to predict survivability of cancer based on relationships between variables from a cancer expert, the training data including previous patient values for the variables, the variables including the survivability;

applying, with a processor, current patient values of the variables for a current cancer patient to the graphic model, the graphic model configured to predict even with one of the variables not having a current patient value as a function of the relationships;

displaying a representation of the graphic model, the representation showing the variables and the relationships remaining after training; and

displaying the survivability for the current cancer patient predicted by the graphic model.

17. The method of claim 16 wherein the cancer comprises non-small cell lung cancer, and wherein the variables comprise tumor load, T-stage, N-stage, number of lymph node stations, who performance, and the survivability.

18. The method of claim 16 wherein the graphic model comprises a Bayesian network, and wherein the relationships include links between nodes for the variables.

19. The method of claim 16 wherein the graphic model configured to predict even with one of the variables not having a current patient value comprises using a value associated with a variable in a position of stronger evidence.

20. The method of claim 16 wherein the graphic model configured to predict even with one of the variables not having a current patient value comprises using a substitute value for the one of the variables, the substitute value selected from a distribution learned from the training data.

21. The method of claim 16 further comprising outputting confidence information with the survivability.