Decision tree models are advantageous when attempting to describe the overall effect of complex decision making, which otherwise would be difficult to describe in the traditional sense. It also allows clinicians to quickly gauge the theoretical impact of altering certain decisions without having to trial those decisions in real time. Thus, it allows health care administrators to implement the most efficient system for the evaluation of complex medical problems.
The decision regarding which diagnostic test to employ in NHGIB patients can become quite complex and involve multiple factors. The goal for creating the current decision tree model was to describe the effect (diagnostic costs and relative times) of traditional diagnostic decisions made by clinicians for patients presenting to the ED with NHGIB. Our study is the first to use a decision tree to analyze the incurred diagnostic costs and relative times associated with using a primary diagnostic VCE, EGD, or COLO in patients presenting with NHGIB. We found VCE, when used as the primary diagnostic procedure in patients admitted to the hospital floor with NHGIB, lead to quicker diagnostic times and less incurred cost when compared to a primary diagnostic EGD or COLO.
In our model, we found the most common primary diagnostic procedure chosen by clinicians was EGD followed by COLO and VCE in both the floor group and ICU group. This aligns well with traditional diagnostic methodology as most patients presenting with NHGIB symptoms (melena, guaiac + stool with anemia), will undergo an EGD first, followed by a COLO if the EGD is non-diagnostic, and finally a VCE if both are non-diagnostic (7–12). However, for patients admitted to the floor, we found the diagnostic probability for a primary EGD and primary COLO was significantly lower than VCE. As a consequence, we noted increasing observed diagnostic costs in floor patients undergoing primary EGDs and COLOs as a significant portion of these patients underwent multiple subsequent non-diagnostic procedures. In contrast, VCE maintains a relatively higher diagnostic yield and consequently accumulates overall lower costs.
The low diagnostic yields of COLOs and EGDs correlate well anecdotally and are likely due to the changing epidemiology of GIB lesions. These findings were corroborated in our recent retrospective analysis demonstrating similar low diagnostic yields with significant variation in location of culprit lesions (35). Peptic ulcers, albeit still common, are becoming less prevalent and NHGIB symptoms such as melena, guaiac + stool with anemia, and hematochezia are less localizing (12, 15–17). Hence, a primary EGD in NHGIB patients may miss a small bowel or right colonic lesion and by the time a COLO or VCE is performed the bleeding lesion has stopped. After all, given enough time most bleeding lesions will resolve spontaneously (29). The diagnostic lag time as well as the need for proper prep time, may also explain the low diagnostic yields of COLOs observed in our model when used as a primary or secondary diagnostic procedure for NHGIB patients admitted to the floor (18, 27–28).
In a similar fashion, our model demonstrated lower accumulated diagnostic times for floor patients undergoing a primary VCE versus an EGD or COLO. This, once again, is a direct consequence to the time accumulated completing secondary and tertiary diagnostic tests after a large number of non diagnostic primary EGDs and COLOs. More specifically, whereas a positive VCE lends itself to only one possible subsequent procedure (only done for further clarification and treatment of the lesion rather than an initial diagnosis), a non-diagnostic EGD or COLO allows for multiple options for diagnostic testing and the array of possibilities leads to an accumulation in time.
Interestingly, similar results were not seen in patients admitted to the ICU as EGDs and COLOs incurred slightly less costs and times than VCE. According to our model, the similarities in cost likely stems from the increased diagnostic probabilities of all three procedures which seems to correlate well with quicker times to the first diagnostic procedure. Interestingly, the average time for a primary diagnostic VCE in the ICU group was 8 hours and all 5 of these patients obtained a diagnosis. This correlates well with previous studies showing the faster VCE is done from the time of index bleeding the higher the likelihood of finding the source of bleeding (31, 32).
Sensitivity analysis done in our model suggests the cost utility of EGDs and COLOs would be equivalent to VCEs if the diagnostic probabilities were substantially increased (70% for EGD and 80% for COLO) for patients admitted to the floor. Although EGD has great diagnostic yields for true UGI bleeding lesions (30), our experience and results of our study indicate clinicians are having difficult times accurately predicting location of bleeding for patients presenting with NHGIB (12, 35). Moreover, given the inherent variability in bleeding lesions throughout the GI tract and an inability to visualize the GI tract past the duodenum, it seems unrealistic to achieve a diagnostic yield greater than the 70% required for EGDs to become diagnostically more cost efficient. Similarly, bowel prep time and lack of rapid visualization of the colon, would make it difficult to achieve a high diagnostic probability for COLOs, unless done urgently (18, 27–28). VCE on the other hand has a high potential to rapidly detect location of bleeding throughout most of the GI tract. While VCE has traditionally been used primarily for suspected small bowel lesions and rarely as a primary diagnostic choice for patients presenting with GIB (19–21), pilot studies using VCE as a primary diagnostic tool for UGIB have demonstrated comparable results to EGD (22–25). More recently, we performed a randomized controlled trial comparing early VCE use to the traditional diagnostic methodology in the work up of NHGIB. We found VCE was able to determine the location of bleeding significantly faster and with better accuracy then EGD and COLO alone (36). It is then to no surprise that the diagnostic probability for VCE in NHGIB floor patients was substantially higher than EGDs or COLOs in our model, and consequently incurs less overall costs and times.
Although VCE has the ability to determine the location of bleeding through most of the GI tract with one test, at times it may have difficulty clarifying the specific cause for bleeding. Thus, secondary diagnostic/therapeutic procedures are usually required for further clarification. Critics will suggest that this will add additional overall cost and may decrease VCE cost utility. However, our sensitivity analysis suggests that even if the diagnostic probability of VCE were to reach 100%, meaning a significant increase in subsequent procedures, the total incurred costs will still be at most equivalent to that of a primary EGD for patients admitted to the floor. The benefit of VCE seems to not only stem from its inherent ability to visualize most of the GI tract with one test but also from its ability to rapidly visualize most of the GI tract and detect intermittently bleeding vascular lesions, which was readily seen in our aforementioned RCT (36).
Only one previous study (Meltzer et al) has designed a decision tree to evaluate the cost effectiveness of VCE compared to other non-invasive diagnostic methods with similar results to ours, but it only looked at patients with acute UGIB and used inferred probabilities (26). Our study supplants Meltzers by using observed probabilities and costs, rather than calculated assumptions, to depict how low diagnostic yields of a primary EGD and COLO can potentially increase overall cost and diagnostic times in the difficult to diagnose NHGIB patient admitted to the floor. In addition, the results of our study indicate that current diagnostic methodologies have room for improvement and exploiting VCEs rapid diagnostic capabilities may be a potential means to curtail unnecessary GIB associated health care costs. We suggest rapid deployment of VCE in the ED setting may increase overall diagnostic probabilities with more directed therapeutic modalities (EGD, COLO, SBE), not only leading to improved diagnostic times but also preventing unnecessary costly hospital admissions. Clearly, prospective studies using a similar decision tree analysis are required to confirm this hypothesis, but the preliminary results seem promising.
There are several limitations of the study beginning with the retrospective nature of the study. The decision tree model was designed to retrospectively follow and analyze the clinician’s decision to perform certain procedures. Thus, assumptions were made that clinicians were using traditional modalities and clinical parameters to determine which procedures should be performed first. Given the inherent variability in clinical judgments and the ambiguity in these clinical parameters, assumptions made based off the decision tree analysis may be oversimplified. Moreover, costs calculated only depicted the cost associated with admitting a patient and their respective diagnostic course. Practically, there are significantly more cost incurred due to other modalities, including medications use, blood transfusions, imaging, therapeutic procedures, etc., that were not included. Thus, costs depicted in our model were not directed at showing actual monetary costs associated with clinical decisions but merely a relative comparison between the three main diagnostic procedures used to diagnose NHGIB. In extension, the total times to diagnosis cannot be interpreted as actual hours to diagnosis since the decision tree is designed to incorporate probabilities with unit measures and is generally additive. However, comparing the total unit hours accumulated between all three procedures allows for at least a relative depiction of which procedures may be quicker at obtaining a diagnosis with respect to each other.
In conclusion, our current decision tree model for NHGIB patients demonstrates a potential opportunity to curtail costs associated with managing patients presenting with NHGIB. In addition, it provides clinicians a means to measure effects of changing certain variables before implementation in real time. Our model, along with other recently published data, suggest early use of VCE may be a more cost-effective method to identify culprit gastrointestinal lesions in NHGIB patients. Further prospective studies are needed to corroborate these findings.