1. Hazlina Hamdan 31 March 2009 1 Modelling survival prediction in medical data By Hazlina Hamdan Dr. Jon Garibaldi

2. Hazlina Hamdan 31 March 2009 2 Presentation Content  Introduction Cancer and Breast Cancer Medical Prognosis Survival Analysis  Research Background Aims & Objectives  Understanding Previous Approach Artificial Neural Network PLANN  Analysis and Results  Conclusion  Future Work

3. Hazlina Hamdan 31 March 2009 3 Cancer  Cancer is basically a disease which occurs when cells behave abnormally and divide out of control – form visible mass or tumour.  There are two general types of tumours namely: benign malignant  Breast cancer is the most common cancer causing fatality amongst women.  Breast cancer is a malignant tumour that develops from the uncontrolled growth of cells in the breast.

4. Hazlina Hamdan 31 March 2009 4 Medical Prognosis  The principal factor in estimating of cure, complication, disease recurrence or survival for a patient or group of patients after treatment.  Prognosis is important because the type and intensity of the medications are based on it.  Prognosis is only a prediction.

5. Hazlina Hamdan 31 March 2009 5 Survival Analysis  The analysis of data that corresponds to the time from when an individual enter a study until the occurrence of some particular event or end-point.  Concerned with the comparison of survival curves for different combinations of risk factors.  Data contains uncensored (reach until end point) and censored (lost to follow-up or die from unrelated cause) observations.

6. Hazlina Hamdan 31 March 2009 6 Survival Analysis – Survival Function  Probability an individual survive at least up to a certain time t. S(t l )=P(T≥t)  Kaplan-Meier survival curve.

7. Hazlina Hamdan 31 March 2009 7  Probability an individual will die at a certain time, conditioned on survival up to that time, and denotes the instantaneous death rate. (Collet D., 1994) h l = P(T Є A l |T>t l-1 ) = f l /S(t l-1 ) known also as conditional failure probability  Survival and Hazard function are related to each other S(t)=∏(1-h l ) l:t l ≤t Survival Analysis – Hazard Function

8. Hazlina Hamdan 31 March 2009 8 Research Objectives  Understand previous approaches.  Apply previous approaches to our data.  Develop novel approaches based on Artificial Neural Network (ANN) and Fuzzy method.  In clinical perspective is to assist doctor in predicting survival of individual patients and planning future treatments.

9. Hazlina Hamdan 31 March 2009 9 Previous Approach Artificial Neural Network (ANN)  Artificial Neural Network (ANN) is defined as an information processing system inspired by the structure of the human brain.  ANN gathers its knowledge by detecting a common pattern and relationships in raw data, then learning from such relationships and adapting the results as required.  The knowledge is then used to predict the outcome for new combinations of data.

10. Hazlina Hamdan 31 March 2009 10 Previous Approach - ANN  In the field of medicine, ANN have been used since the late 1980s, initially as an aid to diagnosis and treatment, and recently as a tool for the analysis of survival data in the presence of censorship.  The ability of neural networks to generalise to new cases based on existing patterns is used as a basis to compute and predict the survival of individual patient or group of patients.

11. Hazlina Hamdan 31 March 2009 11 Previous Approach-Feed-Forward ANN Transfer Function Variables 1…x Inputs Hidden units Outputs Patients A A 1 … A x. N N 1 … N x bias........................

12. Hazlina Hamdan 31 March 2009 12 Previous Approaches-PLANN  Partial Logistic Artificial Neural Network  Proposed by Biganzoli et. al (1998)

13. Hazlina Hamdan 31 March 2009 13 PLANN Model

14. Hazlina Hamdan 31 March 2009 14 PLANN Model-Pre-processing  Categorical variables  indicator variables  Continuous variables  range(-1,1) or (0,1) Treatment TypeIndicator variables Radiotherapy100 Hormone therapy010 Chemotherapy001

15. Hazlina Hamdan 31 March 2009 15 PLANN Model-Pre-processing  Training data - each subjects are replicated for all the intervals in which the subjects is observed and coupled with the event indicator TimeSizeTreat1Treat2Treat3Event Subject131.01001 Subject250.50010 TimeSizeTreat1Treat2Treat3Event Subject111.01000 2 1000 3 1001 Subject210.50010 2 0010 3 0010 4 0010 5 0010

16. Hazlina Hamdan 31 March 2009 16 PLANN Model-Pre-processing  Testing – each subjects are replicated into full number of time interval of observed with all event indicator as zero. TimeSizeTreat1Treat2Treat3Event Subject111.01000 2 1000 3 1000 4 1000 5 1000 Subject210.50010 2 0010 3 0010 4 0010 5 0010 TimeSizeTreat1Treat2Treat3Event Subject131.01001 Subject250.50010

17. Hazlina Hamdan 31 March 2009 17 PLANN Model-Post-processing  Predicted hazard is the mean calculated from the distribution of the activation.

18. Hazlina Hamdan 31 March 2009 18 Analysis and Result  Head and Neck Cancer – disease recurrence Radiation therapy (Arm A) Radiation + Chemotherapy (Arm B) Total patients5145 End of time interval (in month) 4776 Total patient recur until end interval 4231 Total patient lost to follow up 914 Total training replication628967 Total testing4776

19. Hazlina Hamdan 31 March 2009 19 Analysis and Result

20. Hazlina Hamdan 31 March 2009 20 Analysis and Result

21. Hazlina Hamdan 31 March 2009 21 Analysis and Result

22. Hazlina Hamdan 31 March 2009 22 Conclusion  ANN have been considered as alternative methods for analysis of survival for individual patient or group of patients.  A smooth discrete hazard possible be model by treating the time interval and the covariates as an input variable with standard feed forward network and logistic activation function.

23. Hazlina Hamdan 31 March 2009 23 Future Work  Implementing PLANN model to our data (breast cancer data from QMC).  Develop fuzzy set rules in producing the survival rate prediction for breast cancer patient.

24. Hazlina Hamdan 31 March 2009 24 References  Bishop, C. M. (1995). Neural Networks for Pattern Recognition, Oxford University Press Inc., New York,.  Burke, H.B., Goodman, P.H., Rosen, D.B., Henson, D.E., Weinstein, J.N., Harrell, F.E., Marks, J.R., Winchester, D.P. & Bostwick, D.G. (1997). Artificial neural network improve the accuracy of cancer survival prediction. Cancer, vol. 79, pp.857-862  Collett, D. (1994). Modelling Survival Data In Medical Research. Chapman and Hall, London.  Elia Biganzoli, P. B. L. M. E. M. (1998). "Feed forward neural networks for the analysis of censored survival data: a partial logistic regression approach." Statistics in Medicine 17(10): 1169-1186.  Lisboa, P. J. G., H. Wong, et al. (2003). "A Bayesian neural network approach for modelling censored data with an application to prognosis after surgery for breast cancer." Artificial Intelligence in Medicine 28(1): 1-25.  Ohno-Machado, L. (2001). "Modeling Medical Prognosis: Survival Analysis Techniques." Journal of Biomedical Informatics 34(6): 428-439.  Ripley, R. M., A. L. Harris, et al. (1998). "Neural network models for breast cancer prognosis." Neural Computing & Applications 7(4): 367-375.  Ravdin, P. and G. Clark (1992). "A practical application of neural network analysis for predicting outcome of individual breast cancer patients." Breast Cancer Research and Treatment 22(3): 285-293.