The use of POTTER (Predictive Optimal Trees in Emergency Surgery Risk) calculator to predict mortality and complications in patients submitted to Emergency Surgery

ABSTRACT Introduction: the ability of the care team to reliably predict postoperative risk is essential for improvements in surgical decision-making, patient and family counseling, and resource allocation in hospitals. The Artificial Intelligence (AI)-powered POTTER (Predictive Optimal Trees in Emergency Surgery Risk) calculator represents a user-friendly interface and has since been downloaded in its iPhone and Android format by thousands of surgeons worldwide. It was originally developed to be used in non-traumatic emergency surgery patients. However, Potter has not been validated outside the US yet. In this study, we aimed to validate the POTTER calculator in a Brazilian academic hospital. Methods: mortality and morbidity were analyzed using the POTTER calculator in both trauma and non-trauma emergency surgery patients submitted to surgical treatment between November 2020 and July 2021. A total of 194 patients were prospectively included in this analysis. Results: regarding the presence of comorbidities, about 20% of the population were diabetics and 30% were smokers. A total of 47.4% of the patients had hypertensive prednisone. After the analysis of the results, we identified an adequate capability to predict 30-day mortality and morbidity for this group of patients. Conclusion: the POTTER calculator presented excellent performance in predicting both morbidity and mortality in the studied population, representing an important tool for surgical teams to define risks, benefits, and outcomes for the emergency surgery population.

When ESS was developed, it was suggested to be a better predictive model for emergency surgery patients [11][12][13] .Nonetheless, all these aforementioned risk stratification models are based on the idea that the variables used to calculate risk interact in a linear and additive manner.However, medical reality suggests that patients' comorbidities and disease markers interact in a complex, non-linear way and that some variables may gain or lose strength depending on the presence or absence of other variables 14 .
In this context, the Artificial Intelligence (AI)powered POTTER (Predictive Optimal Trees in Emergency Surgery Risk) 4 calculator was recently developed using nearly 400 thousand emergency surgery patients and uses a non-linear, novel, and transparent machine learning methodology to estimate the risk of postoperative mortality and complications.The POTTER user-friendly interface has since been downloaded in its However, POTTER has not been validated outside the US yet.In this study, we aimed to validate the POTTER calculator in Brazil not only for emergency surgery patients but also expand its use for trauma patients, aiming to evaluate its capacity to predict the same variables as for emergency surgery cases.

Patient population
This validation study was carried out in a southeastern Brazilian city, Sorocaba, with a population estimated at around one million people.All patients over 18 years old who were admitted to the General Surgery service at Conjunto Hospitalar de Sorocaba, between November 2020 and July 2021, and were submitted to any kind of emergency surgery procedure were included and are presented in table X as per the admission diagnosis according to the medical records.Both trauma and nontrauma emergency surgery patients were included.
Although the POTTER application was developed using artificial intelligence using data from patients undergoing emergency surgery who were not trauma victims, we opted to add trauma cases undergoing surgical treatment to evaluate if the ability to assess the risks of complications and deaths would be similar to cases of patients who were not trauma victims.The IRB review committee at the Conjunto Hospitalar de Sorocaba reviewed and approved this study -IRB register number 5.013.427.as demonstrated in Table 4 and Figures 1 and 2.
The performance of POTTER to predict morbidity was evaluated for both ESTG and ESG according to Tables 5 and Figures 3 and 4, and in all the groups we were able to demonstrate a p<0.001, proving the applicability of the POTTER.

DISCUSSION
In

R E S U M O R E S U M O
iPhone and Android format by thousands of surgeons worldwide.POTTER derivation and validation have been previously described 4 .Briefly, all patients who underwent emergency surgery in the ACS NSQIP database (2007-2013) were used to train Optimal Classification Trees (OCT) for the development and validation of the POTTER calculator.OCTs are novel, interpretable, machine learning (ML)-based methodologies that follow a sequence of splits (nodes) on key variables to make a final prediction.POTTER effectively predicts the postoperative outcomes of emergency surgery patients and outperforms all other risk calculators in the field (the c-statistic for predicting mortality in EGS (Emergency General Surgery) patients is 0.92).
morbidity, which further confirms POTTER's superiority as a predictive model.Our study is the first international validation of this artificial intelligence tool developed in the United States.Furthermore, due to our inclusion of both trauma and non-trauma patients requiring emergent surgical intervention, the results provide evidence of its usefulness and applicability to a broader population than that previously described.Together with its simple and user-friendly interface in the form of a smartphone application, makes POTTER very easy to deploy at the patient's bedside in the emergency department 5 , which makes it particularly useful in settings where access to a comprehensive electronic health record or similar interface is limited, especially in places like public hospitals in developing countries where internet access is still very limited, making this tool practical and accessible.

Table 1 -
Diagnosis at admission for patients submitted to surgical procedure.Data variables and POTTER predictionMedical records were systematically reviewed, and the following information's were collected according to the requirements of the app based on artificial intelligence database to calculate the complications and mortality rates: Age, laboratory values (hematocrit, white blood cell count, platelet, sodium, potassium, blood urea nitrogen, creatinine, albumin, bilirubin, serum glutamicoxaloacetic transaminase, alkaline phosphatase, partial thromboplastin time, international normalized ratio), comorbidities (COPD, diabetes, smoking, hypertension, acute renal failure, ascites, congestive renal failure, cancer bleeding disorders), intensive care unit (ICU) admission, and complications (fistula, septic shock, aponeurosis dehiscence, pulmonary thromboembolism, anastomosis dehiscence, wound infection, intracavitary abscess, and evisceration), The ACS_NSQIP definitions were used in data collection.Using the collected data, the POTTER predictions of 30-day mortality and 30-day morbidity were calculated for each patient using the existing algorithms and phone application.The primary outcomeStatistical analysisThe area under the receiver operator characteristic curve (AUC), or c-statistic measure, was used to assess the relationship between POTTER's predictions and the outcomes of interest.STATA Software, version 15.1 was used for statistical analysis (Stata Corp).about 20% of the population were diabetics and 30% were smokers.A total of 47.4% of the patients were hypertensive pre-admission.

Table 2 -
Demographic characteristics of the population.

Table 4 -
Predictive Performance of POTTER for 30-day Mortality ESTG and ESG groups.

Table 3 -
Pre-operative comorbidities in the study population.

Table 5 -
Predictive Performance of POTTER for 30-day Combined Morbidity ESTG and ESG groups.