RISK OF GASTROINTESTINAL BLEEDING IN PATIENTS WITH LEFT VENTRICULAR ASSIST DEVICE: A NOVEL PREDICTION MODEL PROPOSAL USING SUPERVISED MACHINE LEARNING

[Introduction] Early detection of gastric cancer offers favorable treatment outcomes. Second-generation narrow-band imaging (NBI) is expected to improve the detection of early gastric cancer (EGC); however, the detection rate is not superior to that of white light imaging (WLI). The latest endoscopic system EVIS X1 (Olympus, Tokyo, Japan) includes a higher-definition WLI, third generation (3G)-NBI, and texture and color enhancement imaging (TXI). We conducted a randomized phase II trial to compare the ability of 3G-NBI, TXI, and WLI to detect gastric neoplasms (GN) (jRCT1032210213).
[Methods] The eligibility criteria included patients aged 20–85 years with either of the following: 1) scheduled surveillance endoscopy after endoscopic resection for GN, or endoscopic resection, chemotherapy, or radiotherapy for esophageal cancer, or 2) scheduled preoperative endoscopy for known GN or esophageal cancer. Written informed consent was obtained from all participants. The EVIS X1 system and a high-definition gastroscope with an optical zoom (GIF-XZ1200) were used. Patients were randomly assigned in a 1:1:1 ratio to the 3G-NBI (primary 3G-NBI and secondary WLI), TXI (primary TXI and secondary WLI), and WLI (primary and secondary WLI) arms, with the WLI arm set as the reference arm. Non-magnifying primary observation was performed to detect the GN lesions. After completing the primary observation of the entire stomach, a secondary WLI was immediately performed by the same endoscopist. All suspected GN lesions were biopsied at the end of the examination. Pathological diagnoses were made based on biopsied tissue or specimens obtained from endoscopic or surgical resection by expert pathologists at each institution. The primary endpoint was GN detection rate in the primary observation, including cancer and adenoma. The other endpoints were the miss rate for GN, EGC detection rate, and positive predictive value (PPV) for the diagnosis of GN in the primary observation. We assumed that the primary endpoint would be 3.0% for one image-enhanced endoscopy (IEE) and > 4.3% for the other. The sample size was set to 300 per arm to ensure that 80% or more of the participants correctly selected the most promising IEE.
[Results] 901 patients were enrolled from six institutions and assigned to the 3G-NBI, TXI, and WLI arms (300/300/301). Of these, 222 (24.6%) underwent preoperative examination. The GN detection rate in the 3G-NBI, TXI, and WLI arms were 7.3%, 5.0%, and 5.6%, respectively, with the 3G-NBI showing the highest detection rate. In addition, the miss rates for GN were 1.0%, 0.7%, and 1.0%, the EGC detection rates were 5.7%, 4.0%, and 5.6%, and the PPVs for the diagnosis of GN were 36.5%, 21.3%, and 36.8% in the 3G-NBI, TXI, and WLI arms, respectively.
[Conclusion] 3G-NBI is the most promising modality for the detection of GN when compared with TXI and WLI.

Background & Aims:
Gastrointestinal (GI) bleeding is a significant complication of left ventricular assist device (LVAD) placement and can be life-threatening. If LVAD patients at higher risk of bleeding could be identified, increased efforts to prevent and manage GI bleeding could be instituted. In this study we aimed to derive a GI-bleed risk prediction model in patients with LVAD, by means of supervised machine learning, using a large retrospective data from the University of Utah heart-transplantation database.

Methods:
We utilized prospectively maintained database comprising 491 LVAD patients between 2004-2022. Institutional review board approval was obtained prior to data collection. The penalized Least Absolute Shrinkage and Selection Operator (LASSO) regression was used to select the best predictors of GI bleeding in LVAD patients, using 10-fold cross validation to develop a practical estimation of the predictive performance, by means of the area under the curve (AUC). Bootstrapped Bias corrected 95% confidence intervals (CI) for the AUC were generated. Thirty potential clinical predictors were introduced in the model. The model disagreement between the predicted and the observed outcomes were evaluated using the Brier score and calibration belt.

Results:
The GI bleeding risk in patients with LVAD was 26.06% (n=128) in the current sample. Three out of the 30 predictors were retained by the prediction model. These were: duration of LVAD implantation (<12 months, 13-24 months, >24 months), anticoagulant use, and LVAD as destination therapy. The model's ability to distinguish LVAD patients at high risk of GI-bleed was outstanding based on the AUC=0.9 (95%CI 0.82-0.92) (Figure 1A). There was no evidence of miscalibration for this model (test statistic=4.61; p=0.10), demonstrating good performance (Brier score=0.05).
A score was assigned to each predictor with a maximum score of 6. Destination therapy as an indication for LVAD transplant =1, LVAD implantation duration =1 if the duration was <12 months, 2 if the duration was 13-24 months, and 3 if the duration was >24 months. A score of 2 was assigned to anticoagulant use. The purpose of score derivation from LASSO regression predictors was to facilitate an understanding, applicability, and practicality for the proposed risk prediction model. The AUC of the Score was 0.90 (95% CI 0.87-0.93) and showed no statistical difference from the AUC of the LASSO model (P=0.49) (Figure 1B). The Youden index determined that cut-off score of 3 had a sensitivity of 69.5% and specificity of 99.1% (Table 1).

Conclusions:
The proposed model presents prediction variables in identifying patients at risk of GI bleeding post LVAD placement. Duration of LVAD implantation, anticoagulant use, and LVAD as destination therapy were the top variables retained. Additional studies are warranted to validate our findings externally.