Dados do Trabalho

Título

Proposal of a new algorithm for continuous outcomes and quality analysis in surgery.

Introdução

Quality has many dimensions. It consists of performance, reliability, durability, serviceability, esthetics, features, perceived value, and conformance to standards. These dimensions may be measured and analyzed to provide feedback on quality and allow its improvement.
Surgery is a service to achieve a result, but a feature of surgery is the risk of undesirable results. These may impact the outcome and its occurrence may be predicted by preoperative and intraoperative variables.

Objetivo

Our objective is to propose an algorithm to evaluate outcomes using industry-proven and well-established methods for quality evaluation using the time dimension and associated with machine learning methods. We believe this algorithm may provide insights to improve quality and outcomes for patients.

Métodos

The thyroid dataset is collected in real-time and includes patients submitted to thyroid surgery from January 2014 to December 2021. The oral cancer is retrospectively collected and includes patients treated between January 1985 and December 2020. Surgical complications were within 30-days of surgery. Oncologic outcomes for oral cancer patients were surgical margins, lymph node yield, preoperative staging, proper adjuvant treatment, preoperative anesthetic evaluation, and proper antibiotic use.
We developed classification models considering surgical complications as outcome. For each model, we calculated the accuracy, sensitivity, specificity, and negative and positive predictive values.
All analyses were performed using the R statistical package. Predictive models were developed using logistic regression, support vector machines, decision trees, random forest, and multilearner classifiers. The datasets were divided into development and validation sets.
IRB approval Nr 2018/20B.

Resultados

The thyroid set includes 7,951 patients. PO complications occurred in 1,549 patients (19.4%). The monthly incidence complications incidence ranged from 0% to 75%. In the multilearner combination, voting yielded the best results. It predicted the complication rate for each month, compared it to the actual complication rate, and plotted the difference between them.
The oral cancer set includes 1,918 patients. PO complications occurred in 690 (36.0%), positive margins in 57 (3.0%), inadequate nodal yield in 136 (7.8%), inadequate staging in 214 (11.1%), failure in adjuvancy in 255 (13.3%), unplanned interruptions in 341 (17.8%), lack of anesthetic evaluation in 267 (13.9%) and inadequate antibiotics in 301 (15.7%). In total, 888 patients (46.3%) presented an event. Patients were grouped in batches of 40. The incidence of events ranged from 5% to 57.5%. We fitted logistic regression, decision trees, SVM, RF and multilearner models. We used the RF model to compare observed versus predicted complications rate in each batch.

Conclusões

A temporal or batch evaluation of outcomes and a comparison between predicted versus observed outcomes allows for a timely identification of factors that may be addressed to reduce surgery morbidity and improve outcomes.

Palavras-chave

oral cancer; surgical complications; outcomes analysis