The Clinical Factors associated with upper gastrointestinal hemorrhage

Reference: Freedman, S.B., et al., Predictors of clinically significant upper gastrointestinal hemorrhage among children with hematemesis. J Pediatr Gastroenterol Nutr, 2012. 54(6): p. 737-43.

A study out from Toronto on the clinical risk factors which predict a significant upper gastrointestinal hemorrhage. 

Brief Summary: Hematemesis is the vomiting of blood and there is currently no study which details the clinical findings which predict upper gastrointestinal hemorrhage (UGIH) – defined by bleeding in the upper gastrointestinal tract and which often requires surgery.  The main objective of this retrospective study was to determine the percentage of children with hematemesis who have UGIH and to identify the clinical features which predict UGIH.  A total of 613 eligible children with determined hematemesis from a tertiary care center in Toronto were accrued into the investigation. 

Results:  A total of 27 of the 613 hematemesis patients (4%) had upper gastrointestinal hemorrhage (UGIH).  The clinical features which were deemed to be predictive were: older age (9.7 vs. 2.9 years), vomiting moderate to large amounts of fresh blood, melena, significant medical history, unwell appearance and tachycardia.  Furthermore, children with a medical history of esophageal/gastric varices and a low hemoglobin and platelet count in the blood are also highly indicative of UGIH. 

Implications for Practice:  A patient with hemorrhage and the following clinical findings (older age (9.7 vs. 2.9 years), vomiting moderate to large amounts of fresh blood, melena, significant medical history, unwell appearance and tachycardia) should be wary of UGIH and perform the appropriate diagnostic tests.  If a child presents with high risk features and there is suspicion of UGIH, then a complete blood count is the only test likely to yield clinically helpful information.

Discussion: Good study here. It is very important that Gastroenterologists know these types of clinical risk factor statistics like a baseball manager knows his own players. 

It was good to see that the author’s included the laboratory investigative tests such as hemoglobin and blood count.   You don’t see this in many clinical risk factor studies. 

Commentary on Statistics and Study Design:  The author’s did several things in this study that I really like. I really liked the sample size calculations – good to see that. Also, the analyst checked for normality and used a different test based on whether the data was normal or not.  Also, it was good to see a correction for multiple comparison testing – you rarely (if ever) see this done, even though it should be in most risk factor studies.

                I guess my first major suggestion comes in the form of a presentation standpoint.  The authors essentially broke up the clinical factor results into historical and clinical features (Table 2), medical history (Table 4), and laboratory investigations (Table 5). Personally, I would have combined all these results into one table.  In the table, one could then just have demarcated the different types of clinical features.  It is just a personal preference of mine – it’s fine either way really.

                My other suggestions have more to do with the actual statistical analysis.  The authors did not perform a logistic regression due to the perceived class imbalance of 4% (positive for UGIH) vs. 96% (negative for UGIH) for the response variable. However, I don’t think this should have prevented the use of the logistic regression model. Generally, as long as the class distribution is the same as the total population (which it’s assumed that it would be?), then only having 4% as the positive class would be OK.  You want to read the paper titled “Sampling Bias and Class Imbalance in Maximum-likelihood Logistic Regression” by Freedman et. al. which explains all this well.  Conversely, what the author’s could done is compared the results from the logistic regression to the 2 sample statistical methods that were used (Mann Whitney and 2-sample t-test).  The results should have been roughly the same.  If they were, then it would be OK to just go ahead and use the logistic regression.  In fact, the 2 sample statistical methods would be just as susceptible to a class im-balance as the logistic regression technique if I’m correct – there is really no way around this.

                Also, I see that the author’s excluded blood pressure and oxygen saturation due to missing data with these variables.  This was probably the best approach here, assuming a large percentage of the data was indeed missing. Generally speaking, if you have missing data for greater than 5% of the data for a given variable, then it is best to just remove the variable if possible.

                As the author’s noted, there is a-lot of potential selection and measurement based limitations with this investigation, and I won’t go into all of them – the author’s would certainly be much better at identifying the appropriate variables than I would.  It’s always important that you be as précises as possible with the measurements on both the predictor and response variable.

                A big thanks to the guys up in Toronto for doing this!