During the last meeting, we discussed the fact that some values seem really high.
When the Gleason Score is greater than 8, it appears for few values of Clinical stage and PSA that the chance for the cancer to be organ confined is higher than with a lower Gleason Score.
After another look at Makarov et al. (2007) paper, I noticed that they also had unexpected results when the Gleason Score was greater than 8:
The second change is a previously undescribed phenomenon: a seemingly paradoxical improvement of pathologic stage predictions in men with GS 8 to 10 versus GS 4 + 3. For GS 8 to 10, in every combination of PSA level and clinical stage, the predicted probability of OC was somewhat higher than in GS 4 + 3. A calibration plot for the model demonstrated that the modeled predicted probabilities of OC were very close to observed values (data not shown). In counseling patients, it is important to relay that patients with GS 8 to 10 in this series were carefully selected; 75% of them had been cleared for surgery with negative results on metastatic workup. The predictions from this nomogram do not apply to all patients with high-grade cancer. The Partin tables and other nomograms relating to men with GS 8 to 10 who have undergone RP represent best-case scenarios of a select few, carefully screened patients with
limited, high-grade cancer on biopsy and no evidence of metastasis on radiographic imaging.
Tuesday, 16 June 2009
15/06/09 - ARI Meeting minutes
Meeting date : 15th June 2009 Monday 4:30pm to 6:00pm
Attended by : John, Olivier, Sam
Minutes :
1) Presentation of first nomograms
i. Nomograms using cross tabulation (number of records for each combination of the input variable) and generated using a sampling method without replacement (a record can only be selected once in each sample). A table is generated for each sample and an average nomogram is calculated at the end of the iterations (50) providing also the standard deviation.
The nomogram shows some surprising values. In some cases, the chance of the cancer to be Organ Confined is higher with a Gleason Score greater than 8 than with a Gleason Score of "4+3" or "3+4".
ii. Nomograms using binary logistic regression (closer to what have been done in Makarov's study).
The table seems to match the Partin table even if more work has to be done to reach the same level of accuracy.
2) Notes on logistic regression
Some tests such as R-Square run in SPSS (our statistic tool) shows that the logistic regression is not a method which fits well the data. This would need further study.
We also pointed out that logistic regression predicts cases that have not been recorded in the dataset and can gives high probabilities for such cases (up to 19%). This also raises the question of how well adapted to the problem the logistic regression is.
3) Nomograms to be generated next
i. Using binary logistic regression, use sampling with replacement to smoothen some high values and to get closer to what had been done by Partin.
ii. As Partin actually used a multivariate logistic regression (output can take more than two values), it would be ideal to masterize it and use it on our dataset rather than the binary regression used so far (output can take the values "1" or "0", therefore our pathological stage variable has been divided between four variables, each of them representing a stage).
iii. Partin calibrated his method using a LOESS curve. This should be also run to get a nomogram using the exact same method as Partin. This would allow us to study differences between the UK and US populations on a similar study. Expected figures should not show a huge difference according to Sam.
iv. Finally, a cross tabulation table using the figures of Partin would be interresting to see and to be compared with ours.
4) We have been successfull on the NRP funding application !
5) Next meeting : Monday 20th July 2009, 4:30pm, Ward 44 Seminar Room
Attended by : John, Olivier, Sam
Minutes :
1) Presentation of first nomograms
i. Nomograms using cross tabulation (number of records for each combination of the input variable) and generated using a sampling method without replacement (a record can only be selected once in each sample). A table is generated for each sample and an average nomogram is calculated at the end of the iterations (50) providing also the standard deviation.
The nomogram shows some surprising values. In some cases, the chance of the cancer to be Organ Confined is higher with a Gleason Score greater than 8 than with a Gleason Score of "4+3" or "3+4".
ii. Nomograms using binary logistic regression (closer to what have been done in Makarov's study).
The table seems to match the Partin table even if more work has to be done to reach the same level of accuracy.
2) Notes on logistic regression
Some tests such as R-Square run in SPSS (our statistic tool) shows that the logistic regression is not a method which fits well the data. This would need further study.
We also pointed out that logistic regression predicts cases that have not been recorded in the dataset and can gives high probabilities for such cases (up to 19%). This also raises the question of how well adapted to the problem the logistic regression is.
3) Nomograms to be generated next
i. Using binary logistic regression, use sampling with replacement to smoothen some high values and to get closer to what had been done by Partin.
ii. As Partin actually used a multivariate logistic regression (output can take more than two values), it would be ideal to masterize it and use it on our dataset rather than the binary regression used so far (output can take the values "1" or "0", therefore our pathological stage variable has been divided between four variables, each of them representing a stage).
iii. Partin calibrated his method using a LOESS curve. This should be also run to get a nomogram using the exact same method as Partin. This would allow us to study differences between the UK and US populations on a similar study. Expected figures should not show a huge difference according to Sam.
iv. Finally, a cross tabulation table using the figures of Partin would be interresting to see and to be compared with ours.
4) We have been successfull on the NRP funding application !
5) Next meeting : Monday 20th July 2009, 4:30pm, Ward 44 Seminar Room
Monday, 18 May 2009
Confirmation on data issues
As decided during the last meeting, Sarah Fowler has been contacted in order to confirm our understanding on the BAUS data.
Her reply highlights the fact that a lot of follow-up form have not been returned resulting in many patient records missing in the second table of the database.
From this, we can "officialize" the data distribution and analysis that have been done last week and keep working in this direction.
Her reply highlights the fact that a lot of follow-up form have not been returned resulting in many patient records missing in the second table of the database.
From this, we can "officialize" the data distribution and analysis that have been done last week and keep working in this direction.
First nomogram for pathological stage prediction
Purely based on the data and using SPSS, a first nomogram has been created. It is only based on the number of time each distinct solution appear in the dataset.The results, although being slightly different, seem to reflect the nomograms from the last update of the Partin tables (Makarov et al., 2007) .
We can also notice a high number of 100% probabilities in the table. This essentialy reflects the final size of the dataset.
For instance, for a PSA of 0-2.5 and a Clinical Stage of T1c, no matter the value of the Gleason Score, the table based on the data predicts the pathological stage to be organ confined (2) in any case.
NB: A row with a series of zero values is not displayed in the table
Thursday, 14 May 2009
11/05/09 ARI meeting minutes (by Thomas Lam)
Meeting date : 11th May 2009 Monday 4:30pm to 6:00pm
Attended by : John, Olivier, Tom, Sam
Minutes :
1) Discussed queries that Olivier had in regard to :
i. CT stage : On closer inspection, this actually refers to Clinical T stage rather than staging obtained from CT scan (i.e. DRE stage, as in Partin tables).
ii. Missing or incomplete data : Overall, we have about 1,600 patients with complete data (i.e. pre-op. PSA, Gleason sum score, DRE and post-op. pathological stage) from a total of more than 7,000 on the database. However, there may be around 400 more as Olivier had excluded patients with pT2 disease.
iii. Pre-op. clinical T stage : Will include sub-stage classes of T2a,b, c, etc. If expressed only as 'T2' or 'T3', then should be excluded.
iv. Post-op. pathological stage : Will include only pT2, pT3a, pT3b and Lymph node positive/negative. Hence, all subdivisions of T2 (eg. T2a,b,c) and T2 on its own can be amalgamated as 'T2'. However, pT3 on its own will be excluded as the subdivisions are the main outcome measures.
v. Bott and Emberton paper : We had a close look at this paper. They had almost 2,000 patients but crucially they omitted pre-op. clinical T stage (or DRE); instead they used % core biopsy volume involvement as a surrogate for this. Hence our study will hopefully be more representative of Partin tables than theirs.
2) Time line : We're actually ahead of schedule. Olivier to continue working on the data once we get clarification from Sarah (BAUS) in regard to the database being consistent (in regard to why only half of patients have complete pathological data).
3) Action points : John to contact Sarah in regard to the above.
4) Further grant application : JRI application : result still pending. If we're not successful, John has identified alternative funding sources to carry the work beyond our pilot phase.
5) Next meeting : 15th June 2009, 4:30pm - 5:30pm, Ward 44 Seminar Room (already booked with Lena Thompson). Apologies for my absence as I'll be on annual leave (holiday in sunny Malaysia from 1st-28th June).
Attended by : John, Olivier, Tom, Sam
Minutes :
1) Discussed queries that Olivier had in regard to :
i. CT stage : On closer inspection, this actually refers to Clinical T stage rather than staging obtained from CT scan (i.e. DRE stage, as in Partin tables).
ii. Missing or incomplete data : Overall, we have about 1,600 patients with complete data (i.e. pre-op. PSA, Gleason sum score, DRE and post-op. pathological stage) from a total of more than 7,000 on the database. However, there may be around 400 more as Olivier had excluded patients with pT2 disease.
iii. Pre-op. clinical T stage : Will include sub-stage classes of T2a,b, c, etc. If expressed only as 'T2' or 'T3', then should be excluded.
iv. Post-op. pathological stage : Will include only pT2, pT3a, pT3b and Lymph node positive/negative. Hence, all subdivisions of T2 (eg. T2a,b,c) and T2 on its own can be amalgamated as 'T2'. However, pT3 on its own will be excluded as the subdivisions are the main outcome measures.
v. Bott and Emberton paper : We had a close look at this paper. They had almost 2,000 patients but crucially they omitted pre-op. clinical T stage (or DRE); instead they used % core biopsy volume involvement as a surrogate for this. Hence our study will hopefully be more representative of Partin tables than theirs.
2) Time line : We're actually ahead of schedule. Olivier to continue working on the data once we get clarification from Sarah (BAUS) in regard to the database being consistent (in regard to why only half of patients have complete pathological data).
3) Action points : John to contact Sarah in regard to the above.
4) Further grant application : JRI application : result still pending. If we're not successful, John has identified alternative funding sources to carry the work beyond our pilot phase.
5) Next meeting : 15th June 2009, 4:30pm - 5:30pm, Ward 44 Seminar Room (already booked with Lena Thompson). Apologies for my absence as I'll be on annual leave (holiday in sunny Malaysia from 1st-28th June).
Wednesday, 13 May 2009
Chi Square tests for independance
A series of tests has been ran in order to see if any dependencies between the four variables of our study can be noticed prior to the use of our algorithm.
Simple Chi Square tests show that DRE result and pathological stage, PSA and pathological stage, and Gleason Score and pathological stage are dependent (with p-value=0).
However, when adding an extra level to each test to confirm under circumstances the dependencies previously noticed, it appears that non of the variables are totally dependent anymore. This can be explained by the existence of relations between PSA, DRE result and Gleason Score, for instance, it is more likely that a patient with a low PSA will be at an early clinical stage and have a low Gleason Score.
It is then not possible to conclude dependencies between one of the three input variables and the pathological stage. But the tests strongly confirm that the pathological stage can be predicted from the three input variables when used together.
Simple Chi Square tests show that DRE result and pathological stage, PSA and pathological stage, and Gleason Score and pathological stage are dependent (with p-value=0).
However, when adding an extra level to each test to confirm under circumstances the dependencies previously noticed, it appears that non of the variables are totally dependent anymore. This can be explained by the existence of relations between PSA, DRE result and Gleason Score, for instance, it is more likely that a patient with a low PSA will be at an early clinical stage and have a low Gleason Score.
It is then not possible to conclude dependencies between one of the three input variables and the pathological stage. But the tests strongly confirm that the pathological stage can be predicted from the three input variables when used together.
Tuesday, 12 May 2009
Demographic information of men included in previous versions of the Partin tables and in current project
This table is based on the demographic information table presented in Makarov et al. (2007).
The last column has been added for a better comparison with our current dataset.
Subscribe to:
Comments (Atom)
