Labs Predict IBD Response to Therapy

Algorithms based on common laboratory values outperformed expensive metabolite testing in determining which patients with inflammatory bowel disease are likely to respond to thiopurine therapy, Dr. Akbar K. Waljee and his colleagues reported.

Thiopurines are known to be effective immunomodulators in patients with inflammatory bowel disease (IBD) who have failed 5-aminosalicylic acid therapy. The problem is that thiopurines have a narrow therapeutic index, and individuals vary widely in how they metabolize these agents.

Experienced clinicians can use inexpensive complete blood count and standard blood chemistry values to balance efficacy and risk in individual patients, but this takes expert judgment, and there are no established algorithms.

A more reproducible approach is to measure the metabolites 6-thioguanine (6-TGN) and 6 methylmercaptopurine (6-MMP). Unfortunately, monitoring these metabolites is expensive, and the sensitivity and specificity of this approach are only 62% and 72%, respectively.

In an effort to resolve this dilemma, Dr. Waljee and his colleagues from the University of Michigan, Ann Arbor, used a machine learning technique to tease out the most accurate algorithms based on CBC and blood chemistries (Clin. Gastroenterol. Hepatol. 2010 [doi:10.1016/j.cgh.2009.09.031]).

The investigators used data collected in 774 cases from 346 individuals who were seen at the University of Michigan between May 2004 and August 2006. To be included in the study, the patients had to have had thiopurine metabolite analysis, CBC, and a comprehensive chemistry panel within the same 24-hour period.

Using a randomly selected 70% of the cases, investigators used a statistical technique called the “random forest” method to derive the most accurate algorithms based on data from the CBC and chemistry panels. They then tested that algorithm on the remaining 30% of the cases, comparing the accuracy to that of thiopurine metabolite analysis.

Their primary outcome measure was the area under the receiver operating characteristic curve (AuROC), a standard measure of accuracy.

The random forest algorithm differentiated clinical response from nonresponse with an AuROC of 0.856, compared with 0.594, for 6-TGN levels, a difference that was highly statistically significant.

The most important independent variables in differentiating responders from nonresponders were neutrophil count, alkaline phosphatase, red-cell distribution width, age, and white blood cell count.

The investigators also derived a random forest algorithm that would predict patient nonadherence, and another that would predict which patients were likely to have unfavorable pharmacodynamic responses to thiopurine therapy. Both of those algorithms proved to be significantly better than thiopurine metabolite analysis.

They also developed a simple prediction rule that was reasonably accurate at differentiating responders from nonresponders. Patients with a ratio of mean corpuscular volume (MCV) to white blood cell count (WBC) of 12 or more had a 67% likelihood of having a clinical response, while those with a ratio less than 12 had a 35% likelihood of having a clinical response. This simplified algorithm was significantly worse than the more complex algorithm, but it was still significantly better than metabolite analysis.

The investigators disclosed that the Regents of the University of Michigan, along with several of the study's coauthors, have applied for a patent on the application of machine learning to the prediction of clinical response to thiopurines.

Algorithms based on common laboratory values outperformed expensive metabolite testing in determining which patients with inflammatory bowel disease are likely to respond to thiopurine therapy, Dr. Akbar K. Waljee and his colleagues reported.

Thiopurines are known to be effective immunomodulators in patients with inflammatory bowel disease (IBD) who have failed 5-aminosalicylic acid therapy. The problem is that thiopurines have a narrow therapeutic index, and individuals vary widely in how they metabolize these agents.

Experienced clinicians can use inexpensive complete blood count and standard blood chemistry values to balance efficacy and risk in individual patients, but this takes expert judgment, and there are no established algorithms.

A more reproducible approach is to measure the metabolites 6-thioguanine (6-TGN) and 6 methylmercaptopurine (6-MMP). Unfortunately, monitoring these metabolites is expensive, and the sensitivity and specificity of this approach are only 62% and 72%, respectively.

In an effort to resolve this dilemma, Dr. Waljee and his colleagues from the University of Michigan, Ann Arbor, used a machine learning technique to tease out the most accurate algorithms based on CBC and blood chemistries (Clin. Gastroenterol. Hepatol. 2010 [doi:10.1016/j.cgh.2009.09.031]).

The investigators used data collected in 774 cases from 346 individuals who were seen at the University of Michigan between May 2004 and August 2006. To be included in the study, the patients had to have had thiopurine metabolite analysis, CBC, and a comprehensive chemistry panel within the same 24-hour period.

Using a randomly selected 70% of the cases, investigators used a statistical technique called the “random forest” method to derive the most accurate algorithms based on data from the CBC and chemistry panels. They then tested that algorithm on the remaining 30% of the cases, comparing the accuracy to that of thiopurine metabolite analysis.

Their primary outcome measure was the area under the receiver operating characteristic curve (AuROC), a standard measure of accuracy.

The random forest algorithm differentiated clinical response from nonresponse with an AuROC of 0.856, compared with 0.594, for 6-TGN levels, a difference that was highly statistically significant.

The most important independent variables in differentiating responders from nonresponders were neutrophil count, alkaline phosphatase, red-cell distribution width, age, and white blood cell count.

The investigators also derived a random forest algorithm that would predict patient nonadherence, and another that would predict which patients were likely to have unfavorable pharmacodynamic responses to thiopurine therapy. Both of those algorithms proved to be significantly better than thiopurine metabolite analysis.

They also developed a simple prediction rule that was reasonably accurate at differentiating responders from nonresponders. Patients with a ratio of mean corpuscular volume (MCV) to white blood cell count (WBC) of 12 or more had a 67% likelihood of having a clinical response, while those with a ratio less than 12 had a 35% likelihood of having a clinical response. This simplified algorithm was significantly worse than the more complex algorithm, but it was still significantly better than metabolite analysis.

The investigators disclosed that the Regents of the University of Michigan, along with several of the study's coauthors, have applied for a patent on the application of machine learning to the prediction of clinical response to thiopurines.

Algorithms based on common laboratory values outperformed expensive metabolite testing in determining which patients with inflammatory bowel disease are likely to respond to thiopurine therapy, Dr. Akbar K. Waljee and his colleagues reported.

Thiopurines are known to be effective immunomodulators in patients with inflammatory bowel disease (IBD) who have failed 5-aminosalicylic acid therapy. The problem is that thiopurines have a narrow therapeutic index, and individuals vary widely in how they metabolize these agents.

Experienced clinicians can use inexpensive complete blood count and standard blood chemistry values to balance efficacy and risk in individual patients, but this takes expert judgment, and there are no established algorithms.

A more reproducible approach is to measure the metabolites 6-thioguanine (6-TGN) and 6 methylmercaptopurine (6-MMP). Unfortunately, monitoring these metabolites is expensive, and the sensitivity and specificity of this approach are only 62% and 72%, respectively.

In an effort to resolve this dilemma, Dr. Waljee and his colleagues from the University of Michigan, Ann Arbor, used a machine learning technique to tease out the most accurate algorithms based on CBC and blood chemistries (Clin. Gastroenterol. Hepatol. 2010 [doi:10.1016/j.cgh.2009.09.031]).

The investigators used data collected in 774 cases from 346 individuals who were seen at the University of Michigan between May 2004 and August 2006. To be included in the study, the patients had to have had thiopurine metabolite analysis, CBC, and a comprehensive chemistry panel within the same 24-hour period.

Using a randomly selected 70% of the cases, investigators used a statistical technique called the “random forest” method to derive the most accurate algorithms based on data from the CBC and chemistry panels. They then tested that algorithm on the remaining 30% of the cases, comparing the accuracy to that of thiopurine metabolite analysis.

Their primary outcome measure was the area under the receiver operating characteristic curve (AuROC), a standard measure of accuracy.

The random forest algorithm differentiated clinical response from nonresponse with an AuROC of 0.856, compared with 0.594, for 6-TGN levels, a difference that was highly statistically significant.

The most important independent variables in differentiating responders from nonresponders were neutrophil count, alkaline phosphatase, red-cell distribution width, age, and white blood cell count.

The investigators also derived a random forest algorithm that would predict patient nonadherence, and another that would predict which patients were likely to have unfavorable pharmacodynamic responses to thiopurine therapy. Both of those algorithms proved to be significantly better than thiopurine metabolite analysis.

They also developed a simple prediction rule that was reasonably accurate at differentiating responders from nonresponders. Patients with a ratio of mean corpuscular volume (MCV) to white blood cell count (WBC) of 12 or more had a 67% likelihood of having a clinical response, while those with a ratio less than 12 had a 35% likelihood of having a clinical response. This simplified algorithm was significantly worse than the more complex algorithm, but it was still significantly better than metabolite analysis.

The investigators disclosed that the Regents of the University of Michigan, along with several of the study's coauthors, have applied for a patent on the application of machine learning to the prediction of clinical response to thiopurines.