Machine learning in liver surgery: Benefits and pitfalls

The application of machine learning (ML) algorithms in various fields of hepatology is an issue of interest. However, we must be cautious with the results. In this letter, based on a published ML prediction model for acute kidney injury after liver surgery, we discuss some limitations of ML models and how they may be addressed in the future. Although the future faces significant challenges, it also holds a great potential.

Machine perfusion and the prevention of ischemic type biliary lesions following liver transplant: What is the evidence?

The widespread uptake of different machine perfusion (MP) strategies for liver transplant has been driven by an effort to minimize graft injury. Damage to the cholangiocytes during the liver donation, preservation, or early posttransplant period may result in stricturing of the biliary tree and inadequate biliary drainage. This problem continues to trouble clinicians, and may have catastrophic consequences for the graft and patient. Ischemic injury, as a result of compromised hepatic artery flow, is a well-known cause of biliary strictures, sepsis, and graft failure. However, very similar lesions can appear with a patent hepatic artery and these are known as ischemic type biliary lesions (ITBL) that are attributed to microcirculatory dysfunction rather than main hepatic arterial compromise. Both the warm and cold ischemic period duration appear to influence the onset of ITBL. All of the commonly used MP techniques deliver oxygen to the graft cells, and therefore may minimize the cholangiocyte injury and subsequently reduce the incidence of ITBL. As clinical experience and published evidence grows for these modalities, the impact they have on ITBL rates is important to consider. In this review, the evidence for the three commonly used MP strategies (abdominal normothermic regional perfusion [A-NRP], hypothermic oxygenated perfusion [HOPE], and normothermic machine perfusion [NMP] for ITBL prevention has been critically reviewed. Inconsistencies with ITBL definitions used in trials, coupled with variations in techniques of MP, make interpretation challenging. Overall, the evidence suggests that both HOPE and A-NRP prevent ITBL in donated after circulatory death grafts compared to cold storage. The evidence for ITBL prevention in donor after brain death grafts with any MP technique is weak.

Fast-track protocols in laparoscopic liver surgery: Applicability and correlation with difficulty scoring systems

BACKGROUND

Few series have reported the utility of fast-track protocols (FTP) in minimally invasive liver surgery.

AIM

To report the applicability of FTP in minimally invasive liver surgery and to correlate with difficulty scores.

METHODS

The series of patients undergoing minimally invasive liver surgery from 2014 was analyzed. Iwate, Southampton and Gayet’s scores were compared as predictors of FTP adherence. Accomplishment of FTP was considered within 24-h, 48-h and 72-h. Multivariate models were performed to define discharge < 24 h, < 72 h, complications and readmissions.

RESULTS

From 160 cases, 78 were candidates for FTP, of which 22 (28.2%), 19 (24.4%) and 14 (17.9%) were discharged in < 24-h, 48-h and 72-h, respectively (total = 71.5%). Iwate, Southampton and Gayet’s scores achieved area under the receiver operating characteristic values for < 24-h stay of 0.780, 0.687 and 0.698, respectively. Sensitivity and specificity values for the best score (Iwate) were 87.7% and 66.7%, respectively (cutoff = 5.5). In multivariate models, < 72 h stay and complications revealed body mass index as a risk factor independent from difficulty scores.

CONCLUSION

The development of aggressive FTP is feasible and < 24-h stay can be achieved even in moderate and advanced complexity cases. Difficulty scores, including body mass index value, may be useful to predict which cases may adhere to these protocols.

Short-term outcomes of robotic liver resection: An initial single-institution experience

BACKGROUND

Liver surgery has traditionally been characterized by the complexity of its procedures and potentially high rates of morbidity and mortality in inexperienced hands. The robotic approach has gradually been introduced in liver surgery and has increased notably in recent years. However, few centers currently perform robotic liver surgery and experiences in robot-assisted surgical procedures continue to be limited compared to the laparoscopic approach.

AIM

To analyze the outcomes and feasibility of an initial robotic liver program implemented in an experienced laparoscopic hepatobiliary center.

METHODS

A total of forty consecutive patients underwent robotic liver resection (da Vinci Xi, intuitive.com, United States) between June 2019 and January 2021. Patients were prospectively followed and retrospectively reviewed. Clinicopathological characteristics and perioperative and short-term outcomes were analyzed. Data are expressed as mean and standard deviation. The study was approved by the Institutional Review Board.

RESULTS

The mean age of patients was 59.55 years, of which 18 (45%) were female. The mean body mass index was 29.41 kg/m². Nine patients (22.5%) were cirrhotic. Patients were divided by type of resection as follows: Ten segmentectomies, three wedge resections, ten left lateral sectionectomies, six bisegmentectomies (two V-VI bisegmentectomies and four IVb-V bisegmentectomies), two right anterior sectionectomies, five left hepatectomies and two right hepatectomies. Malignant lesions occurred in twenty-nine (72.5%) of the patients. The mean operative time was 258.11 min and two patients were transfused intraoperatively (5%). Inflow occlusion was used in thirty cases (75%) and the mean total clamping time was 32.62 min. There was a single conversion due to uncontrollable hemorrhage. Major postoperative complications (Clavien–Dindo > IIIb) occurred in three patients (7.5%) and mortality in one (2.5%). No patient required readmission to the hospital. The mean hospital stay was 5.6 d.

CONCLUSION

Although robotic hepatectomy is a safe and feasible procedure with favorable short-term outcomes, it involves a demanding learning curve that requires a high level of training, skill and dexterity.

Machine-learning algorithms for predicting results in liver transplantation: the problem of donor-recipient matching

PURPOSE OF REVIEW

Classifiers based on artificial intelligence can be useful to solve decision problems related to the inclusion or removal of possible liver transplant candidates, and assisting in the heterogeneous field of donor-recipient (D-R) matching.

RECENT FINDINGS

Artificial intelligence models can show a great advantage by being able to handle a multitude of variables, be objective and help in cases of similar probabilities. In the field of liver transplantation, the most commonly used classifiers have been artificial neural networks (ANNs) and random forest classifiers. ANNs are excellent tools for finding patterns which are far too complex for a clinician and are capable of generating near-perfect predictions on the data on which they are fit, yielding excellent prediction capabilities reaching 95% for 3 months graft survival. On the other hand, RF can overcome ANNs in some of their limitations, mainly because of the lack of information on the variables they provide. Random forest algorithms may allow for improved confidence with the use of marginal organs and better outcome after transplantation.

SUMMARY

ANNs and random forest can handle a multitude of structured and unstructured parameters, and establish non explicit relationships among risk factors of clinical relevance.

Use of Peritoneovenous Shunt for the Management of Refractory Ascites

BACKGROUND

Guidelines for the management of refractory ascites (RA) recommend transjugular intrahepatic portosystemic shunting (TIPS), diuretics, and paracentesis as the main strategies, discouraging use of surgical peritoneovenous shunts (PVSs). However, PVSs, including both Denver (DS) or saphenoperitoneal (SPS) modalities, may still have indications. Herein we report our experience with PVSs in the context of modern surgical and anesthetic management.

METHODS

In our unit, PVSs are offered to patients with ascites refractory to diuretics in which TIPS are contraindicated. Heart function and spontaneous bacterial peritonitis must be assessed before surgical indication.

RESULTS

Seven procedures were performed on 5 patients (6-DS, 1-SPS) in 2013. Their mean age was 61 (range, 54-68) years. In 3 patients, the indication was RA without options for liver transplant; 2 patients were on the waiting list for liver transplantation, which were performed to improve renal function and quality of life (QOL). The median hospital stay was 6.5 (range, 3-12) days. All patients were alive after 12 months. One patient died 2 years after the first DS and another later died due to liver insufficiency with patency of the DS. The ascites was well-controlled in 4 of 5 patients at up to 48 months of follow-up. Decreases in diuretics doses, proper weight maintenance, and a dramatic improvement in QOL (measured by a modified Ascites Symptom Inventory-7 [ASI-7] test) were observed after the procedures.

CONCLUSION

PVSs are useful for the treatment of patients with RA who develop resistance to common therapies, leading to a major improvement in QOL. These surgical procedures should be included in the armamentarium of experienced liver surgeons.

Validation of artificial neural networks as a methodology for donor-recipient matching for liver transplantation

In 2014, we reported a model for donor-recipient (D-R) matching in liver transplantation (LT) based on artificial neural networks (ANNs) from a Spanish multicenter study (Model for Allocation of Donor and Recipient in España [MADR-E]). The aim is to test the ANN-based methodology in a different European health care system in order to validate it. An ANN model was designed using a cohort of patients from King's College Hospital (KCH; n = 822). The ANN was trained and tested using KCH pairs for both 3- and 12-month survival models. End points were probability of graft survival (correct classification rate [CCR]) and nonsurvival (minimum sensitivity [MS]). The final model is a rule-based system for facilitating the decision about the most appropriate D-R matching. Models designed for KCH had excellent prediction capabilities for both 3 months (CCR-area under the curve [AUC] = 0.94; MS-AUC = 0.94) and 12 months (CCR-AUC = 0.78; MS-AUC = 0.82), almost 15% higher than the best obtained by other known scores such as Model for End-Stage Liver Disease and balance of risk. Moreover, these results improve the previously reported ones in the multicentric MADR-E database. In conclusion, the use of ANN for D-R matching in LT in other health care systems achieved excellent prediction capabilities supporting the validation of these tools. It should be considered as the most advanced, objective, and useful tool to date for the management of waiting lists. Liver Transplantation 24 192-203 2018 AASLD.