Usability, Acceptability, and Implementation of Artificial Intelligence (AI) and Machine Learning (ML) Techniques in Surgical Coaching and Training: A Scoping Review.

OBJECTIVE: To define the current state of peer-reviewed literature demonstrating the usability, acceptability, and implementation of artificial intelligence (AI) and machine learning (ML) techniques in surgical coaching and training. DESIGN: We conducted a literature search with defined inclusion and exclusion criteria. We searched five scholarly databases: MEDLINE via PubMed, Embase via Elsevier, Scopus via Elsevier, Cochrane Central Register of Controlled Trials, and the Healthcare Administration Database via ProQuest. We followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews (PRISMA-ScR) guidelines. RESULTS: Only 4 articles met the inclusion criteria and used standardized methods for performance evaluation with expert observation. We found no literature examining the impact on performance, user acceptance, or implementation of AI/ML techniques used for surgical coaching and training. We highlight the need for qualitative and quantitative research demonstrating these techniques' effectiveness before broad implementation. CONCLUSION AND RELEVANCE: We emphasize the need for research to specifically evaluate performance, impact, user acceptance, and implementation of AI/ML techniques. Incorporating these facets of research when developing AI/ML techniques for surgical training is crucial to ensure emerging technology meets user needs without increasing cognitive burden or frustrating users.

Linking gene expression to clinical outcomes in pediatric Crohn's disease using machine learning.

Pediatric Crohn's disease (CD) is characterized by a severe disease course with frequent complications. We sought to apply machine learning-based models to predict risk of developing future complications in pediatric CD using ileal and colonic gene expression. Gene expression data was generated from 101 formalin-fixed, paraffin-embedded (FFPE) ileal and colonic biopsies obtained from treatment-naïve CD patients and controls. Clinical outcomes including development of strictures or fistulas and progression to surgery were analyzed using differential expression and modeled using machine learning. Differential expression analysis revealed downregulation of pathways related to inflammation and extra-cellular matrix production in patients with strictures. Machine learning-based models were able to incorporate colonic gene expression and clinical characteristics to predict outcomes with high accuracy. Models showed an area under the receiver operating characteristic curve (AUROC) of 0.84 for strictures, 0.83 for remission, and 0.75 for surgery. Genes with potential prognostic importance for strictures (REG1A, MMP3, and DUOX2) were not identified in single gene differential analysis but were found to have strong contributions to predictive models. Our findings in FFPE tissue support the importance of colonic gene expression and the potential for machine learning-based models in predicting outcomes for pediatric CD.

Factors Associated With Performing IPAA After Total Colectomy for Ulcerative Colitis.

BACKGROUND: IPAA is considered the procedure of choice for restorative surgery after total colectomy for ulcerative colitis. Previous studies have examined the rate of IPAA within individual states but not at the national level in the United States. OBJECTIVE: This study aimed to assess the rate of IPAA after total colectomy for ulcerative colitis in a national population and identify factors associated with IPAA. DESIGN: This was a retrospective cohort study. SETTINGS: This study was performed in the United States. PATIENTS: Patients who were aged 18 years or older and who underwent total colectomy between 2009 and 2019 for a diagnosis of ulcerative colitis were identified within a commercial database. This database excluded patients with public insurance, including all patients older than 65 years with Medicare. MAIN OUTCOME MEASURES: The primary outcome was IPAA. Multivariable logistic regression was used to assess the association between covariates and the likelihood of undergoing IPAA. RESULTS: In total, 2816 patients were included, of whom 1414 (50.2%) underwent IPAA, 928 (33.0%) underwent no further surgery, and 474 (16.8%) underwent proctectomy with end ileostomy. Younger age, lower comorbidities, elective case, and laparoscopic approach in the initial colectomy were significantly associated with IPAA but socioeconomic status was not. LIMITATIONS: This retrospective study included only patients with commercial insurance. CONCLUSIONS: A total of 50.2% of patients who had total colectomy for ulcerative colitis underwent IPAA, and younger age, lower comorbidities, and elective cases are associated with a higher rate of IPAA placement. This study emphasizes the importance of ensuring follow-up with colorectal surgeons to provide the option of restorative surgery, especially for patients undergoing urgent or emergent colectomies. See Video Abstract . FACTORES ASOCIADOS CON LA REALIZACIN DE ANASTOMOSIS ANALBOLSA ILEAL DESPUS DE UNA COLECTOMA TOTAL POR COLITIS ULCEROSA: ANTECEDENTES:La anastomosis ileo-anal se considera el procedimiento de elección para la cirugía reparadora tras la colectomía total por colitis ulcerosa. Estudios previos han examinado la tasa de anastomosis ileo-anal dentro de los estados individuales, pero no a nivel nacional en los Estados Unidos.OBJETIVO:Evaluar la tasa de anastomosis bolsa ileal-anal después de la colectomía total para la colitis ulcerosa en una población nacional e identificar los factores asociados con la anastomosis bolsa ileal-anal.DISEÑO:Se trata de un estudio de cohortes retrospectivo.LUGAR:Este estudio se realizó en los Estados Unidos.PACIENTES:Los pacientes que tenían ≥18 años de edad que se sometieron a colectomía total entre 2009 y 2019 para un diagnóstico de colitis ulcerosa fueron identificados dentro de una base de datos comercial. Esta base de datos excluyó a los pacientes con seguro público, incluidos todos los pacientes >65 años con Medicare.MEDIDAS DE RESULTADO PRINCIPALES:El resultado primario fue la anastomosis ileal bolsa-anal. Se utilizó una regresión logística multivariable para evaluar la asociación entre las covariables y la probabilidad de someterse a una anastomosis ileal.RESULTADOS:En total, se incluyeron 2.816 pacientes, de los cuales 1.414 (50,2%) se sometieron a anastomosis ileo-anal, 928 (33,0%) no se sometieron a ninguna otra intervención quirúrgica y 474 (16,8%) se sometieron a proctectomía con ileostomía terminal. La edad más joven, las comorbilidades más bajas, el caso electivo, y el abordaje laparoscópico en la colectomía inicial se asociaron significativamente con la anastomosis ileal bolsa-anal, pero no el estatus socioeconómico.LIMITACIONES:Este estudio retrospectivo incluyó sólo pacientes con seguro comercial.CONCLUSIONES:Un 50,2% de los pacientes se someten a anastomosis ileo-anal y la edad más joven, las comorbilidades más bajas y los casos electivos se asocian con una mayor tasa de colocación de anastomosis ileo-anal. Esto subraya la importancia de asegurar el seguimiento con cirujanos colorrectales para ofrecer la opción de cirugía reparadora, especialmente en pacientes sometidos a colectomías urgentes o emergentes. (Traducción-Dr. Yolanda Colorado ).

Prediction of Pathologic Complete Response for Rectal Cancer Based on Pretreatment Factors Using Machine Learning.

BACKGROUND: Pathologic complete response after neoadjuvant therapy is an important prognostic indicator for locally advanced rectal cancer and may give insights into which patients might be treated nonoperatively in the future. Existing models for predicting pathologic complete response in the pretreatment setting are limited by small data sets and low accuracy. OBJECTIVE: We sought to use machine learning to develop a more generalizable predictive model for pathologic complete response for locally advanced rectal cancer. DESIGN: Patients with locally advanced rectal cancer who underwent neoadjuvant therapy followed by surgical resection were identified in the National Cancer Database from years 2010 to 2019 and were split into training, validation, and test sets. Machine learning techniques included random forest, gradient boosting, and artificial neural network. A logistic regression model was also created. Model performance was assessed using an area under the receiver operating characteristic curve. SETTINGS: This study used a national, multicenter data set. PATIENTS: Patients with locally advanced rectal cancer who underwent neoadjuvant therapy and proctectomy. MAIN OUTCOME MEASURES: Pathologic complete response defined as T0/xN0/x. RESULTS: The data set included 53,684 patients. Pathologic complete response was experienced by 22.9% of patients. Gradient boosting showed the best performance with an area under the receiver operating characteristic curve of 0.777 (95% CI, 0.773-0.781), compared with 0.684 (95% CI, 0.68-0.688) for logistic regression. The strongest predictors of pathologic complete response were no lymphovascular invasion, no perineural invasion, lower CEA, smaller size of tumor, and microsatellite stability. A concise model including the top 5 variables showed preserved performance. LIMITATIONS: The models were not externally validated. CONCLUSIONS: Machine learning techniques can be used to accurately predict pathologic complete response for locally advanced rectal cancer in the pretreatment setting. After fine-tuning a data set including patients treated nonoperatively, these models could help clinicians identify the appropriate candidates for a watch-and-wait strategy. See Video Abstract . EL CNCER DE RECTO BASADA EN FACTORES PREVIOS AL TRATAMIENTO MEDIANTE EL APRENDIZAJE AUTOMTICO: ANTECEDENTES:La respuesta patológica completa después de la terapia neoadyuvante es un indicador pronóstico importante para el cáncer de recto localmente avanzado y puede dar información sobre qué pacientes podrían ser tratados de forma no quirúrgica en el futuro. Los modelos existentes para predecir la respuesta patológica completa en el entorno previo al tratamiento están limitados por conjuntos de datos pequeños y baja precisión.OBJETIVO:Intentamos utilizar el aprendizaje automático para desarrollar un modelo predictivo más generalizable para la respuesta patológica completa para el cáncer de recto localmente avanzado.DISEÑO:Los pacientes con cáncer de recto localmente avanzado que se sometieron a terapia neoadyuvante seguida de resección quirúrgica se identificaron en la Base de Datos Nacional del Cáncer de los años 2010 a 2019 y se dividieron en conjuntos de capacitación, validación y prueba. Las técnicas de aprendizaje automático incluyeron bosque aleatorio, aumento de gradiente y red neuronal artificial. También se creó un modelo de regresión logística. El rendimiento del modelo se evaluó utilizando el área bajo la curva característica operativa del receptor.ÁMBITO:Este estudio utilizó un conjunto de datos nacional multicéntrico.PACIENTES:Pacientes con cáncer de recto localmente avanzado sometidos a terapia neoadyuvante y proctectomía.PRINCIPALES MEDIDAS DE VALORACIÓN:Respuesta patológica completa definida como T0/xN0/x.RESULTADOS:El conjunto de datos incluyó 53.684 pacientes. El 22,9% de los pacientes experimentaron una respuesta patológica completa. El refuerzo de gradiente mostró el mejor rendimiento con un área bajo la curva característica operativa del receptor de 0,777 (IC del 95%: 0,773 - 0,781), en comparación con 0,684 (IC del 95%: 0,68 - 0,688) para la regresión logística. Los predictores más fuertes de respuesta patológica completa fueron la ausencia de invasión linfovascular, la ausencia de invasión perineural, un CEA más bajo, un tamaño más pequeño del tumor y la estabilidad de los microsatélites. Un modelo conciso que incluye las cinco variables principales mostró un rendimiento preservado.LIMITACIONES:Los modelos no fueron validados externamente.CONCLUSIONES:Las técnicas de aprendizaje automático se pueden utilizar para predecir con precisión la respuesta patológica completa para el cáncer de recto localmente avanzado en el entorno previo al tratamiento. Después de realizar ajustes en un conjunto de datos que incluye pacientes tratados de forma no quirúrgica, estos modelos podrían ayudar a los médicos a identificar a los candidatos adecuados para una estrategia de observar y esperar. (Traducción-Dr. Ingrid Melo ).

Combined kinome inhibition states are predictive of cancer cell line sensitivity to kinase inhibitor combination therapies.

Protein kinases are a primary focus in targeted therapy development for cancer, owing to their role as regulators in nearly all areas of cell life. Recent strategies targeting the kinome with combination therapies have shown promise, such as trametinib and dabrafenib in advanced melanoma, but empirical design for less characterized pathways remains a challenge. Computational combination screening is an attractive alternative, allowing in-silico filtering prior to experimental testing of drastically fewer leads, increasing efficiency and effectiveness of drug development pipelines. In this work, we generated combined kinome inhibition states of 40,000 kinase inhibitor combinations from kinobeads-based kinome profiling across 64 doses. We then integrated these with transcriptomics from CCLE to build machine learning models with elastic-net feature selection to predict cell line sensitivity across nine cancer types, with accuracy R2 ∼ 0.75-0.9. We then validated the model by using a PDX-derived TNBC cell line and saw good global accuracy (R2 ∼ 0.7) as well as high accuracy in predicting synergy using four popular metrics (R2 ∼ 0.9). Additionally, the model was able to predict a highly synergistic combination of trametinib and omipalisib for TNBC treatment, which incidentally was recently in phase I clinical trials. Our choice of tree-based models for greater interpretability allowed interrogation of highly predictive kinases in each cancer type, such as the MAPK, CDK, and STK kinases. Overall, these results suggest that kinome inhibition states of kinase inhibitor combinations are strongly predictive of cell line responses and have great potential for integration into computational drug screening pipelines. This approach may facilitate the identification of effective kinase inhibitor combinations and accelerate the development of novel cancer therapies, ultimately improving patient outcomes.

Integrated single-dose kinome profiling data is predictive of cancer cell line sensitivity to kinase inhibitors.

Protein kinase activity forms the backbone of cellular information transfer, acting both individually and as part of a broader network, the kinome. Their central role in signaling leads to kinome dysfunction being a common driver of disease, and in particular cancer, where numerous kinases have been identified as having a causal or modulating role in tumor development and progression. As a result, the development of therapies targeting kinases has rapidly grown, with over 70 kinase inhibitors approved for use in the clinic and over double this number currently in clinical trials. Understanding the relationship between kinase inhibitor treatment and their effects on downstream cellular phenotype is thus of clear importance for understanding treatment mechanisms and streamlining compound screening in therapy development. In this work, we combine two large-scale kinome profiling data sets and use them to link inhibitor-kinome interactions with cell line treatment responses (AUC/IC50). We then built computational models on this data set that achieve a high degree of prediction accuracy (R2 of 0.7 and RMSE of 0.9) and were able to identify a set of well-characterized and understudied kinases that significantly affect cell responses. We further validated these models experimentally by testing predicted effects in breast cancer cell lines and extended the model scope by performing additional validation in patient-derived pancreatic cancer cell lines. Overall, these results demonstrate that broad quantification of kinome inhibition state is highly predictive of downstream cellular phenotypes.

Analysis of Specimen Mammography with Artificial Intelligence to Predict Margin Status.

BACKGROUND: Intraoperative specimen mammography is a valuable tool in breast cancer surgery, providing immediate assessment of margins for a resected tumor. However, the accuracy of specimen mammography in detecting microscopic margin positivity is low. We sought to develop an artificial intelligence model to predict the pathologic margin status of resected breast tumors using specimen mammography. METHODS: A dataset of specimen mammography images matched with pathologic margin status was collected from our institution from 2017 to 2020. The dataset was randomly split into training, validation, and test sets. Specimen mammography models pretrained on radiologic images were developed and compared with models pretrained on nonmedical images. Model performance was assessed using sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC). RESULTS: The dataset included 821 images, and 53% had positive margins. For three out of four model architectures tested, models pretrained on radiologic images outperformed nonmedical models. The highest performing model, InceptionV3, showed sensitivity of 84%, specificity of 42%, and AUROC of 0.71. Model performance was better among patients with invasive cancers, less dense breasts, and non-white race. CONCLUSIONS: This study developed and internally validated artificial intelligence models that predict pathologic margins status for partial mastectomy from specimen mammograms. The models' accuracy compares favorably with published literature on surgeon and radiologist interpretation of specimen mammography. With further development, these models could more precisely guide the extent of resection, potentially improving cosmesis and reducing reoperations.

Combined kinome inhibition states are predictive of cancer cell line sensitivity to kinase inhibitor combination therapies.

Protein kinases are a primary focus in targeted therapy development for cancer, owing to their role as regulators in nearly all areas of cell life. Kinase inhibitors are one of the fastest growing drug classes in oncology, but resistance acquisition to kinase-targeting monotherapies is inevitable due to the dynamic and interconnected nature of the kinome in response to perturbation. Recent strategies targeting the kinome with combination therapies have shown promise, such as the approval of Trametinib and Dabrafenib in advanced melanoma, but similar empirical combination design for less characterized pathways remains a challenge. Computational combination screening is an attractive alternative, allowing in-silico screening prior to in-vitro or in-vivo testing of drastically fewer leads, increasing efficiency and effectiveness of drug development pipelines. In this work, we generate combined kinome inhibition states of 40,000 kinase inhibitor combinations from kinobeads-based kinome profiling across 64 doses. We then integrated these with baseline transcriptomics from CCLE to build robust machine learning models to predict cell line sensitivity from NCI-ALMANAC across nine cancer types, with model accuracy R2 ~ 0.75-0.9 after feature selection using elastic-net regression. We further validated the model's ability to extend to real-world examples by using the best-performing breast cancer model to generate predictions for kinase inhibitor combination sensitivity and synergy in a PDX-derived TNBC cell line and saw reasonable global accuracy in our experimental validation (R2 ~ 0.7) as well as high accuracy in predicting synergy using four popular metrics (R2 ~ 0.9). Additionally, the model was able to predict a highly synergistic combination of Trametinib (MEK inhibitor) and Omipalisib (PI3K inhibitor) for TNBC treatment, which incidentally was recently in phase I clinical trials for TNBC. Our choice of tree-based models over networks for greater interpretability also allowed us to further interrogate which specific kinases were highly predictive of cell sensitivity in each cancer type, and we saw confirmatory strong predictive power in the inhibition of MAPK, CDK, and STK kinases. Overall, these results suggest that kinome inhibition states of kinase inhibitor combinations are strongly predictive of cell line responses and have great potential for integration into computational drug screening pipelines. This approach may facilitate the identification of effective kinase inhibitor combinations and accelerate the development of novel cancer therapies, ultimately improving patient outcomes.

Estimating Risk of Locoregional Failure and Overall Survival in Anal Cancer Following Chemoradiation: A Machine Learning Approach.

BACKGROUND: Optimal treatment of anal squamous cell carcinoma (ASCC) is definitive chemoradiation. Patients with persistent or recurrent disease require abdominoperineal resection (APR). Current models for predicting need for APR and overall survival are limited by low accuracy or small datasets. This study sought to use machine learning (ML) to develop more accurate models for locoregional failure and overall survival for ASCC. METHODS: This study used the National Cancer Database from 2004-2018, divided into training, validation, and test sets. We included patients with stage I-III ASCC who underwent chemoradiation. Our primary outcomes were need for APR and 3-year overall survival. Random forest (RF), gradient boosting (XGB), and neural network (NN) ML-based models were developed and compared with logistic regression (LR). Accuracy was assessed using area under the receiver operating characteristic curve (AUROC). RESULTS: APR was required in 5.3% (1,015/18,978) of patients. XGB performed best with AUROC of 0.813, compared with 0.691 for LR. Tumor size, lymphovascular invasion, and tumor grade showed the strongest influence on model predictions. Mortality was 23.6% (7,988/33,834). AUROC for XGB and LR were similar at 0.766 and 0.748, respectively. For this model, age, radiation dose, sex, and insurance status were the most influential variables. CONCLUSIONS: We developed and internally validated machine learning-based models for predicting outcomes in ASCC and showed higher accuracy versus LR for locoregional failure, but not overall survival. After external validation, these models may assist clinicians with identifying patients with ASCC at high risk of treatment failure.

Predicting readmission after bariatric surgery using machine learning.

BACKGROUND: While bariatric surgery is an effective method for achieving long-term weight loss, postoperative readmissions are associated with negative clinical outcomes and significant costs. OBJECTIVES: We aimed to use machine learning (ML) algorithms to predict readmissions and compare results to logistic regression. SETTING: Hospitals participating in the Metabolic and Bariatric Surgery Accreditation and Quality Improvement Program, United States. METHODS: Patients who underwent sleeve gastrectomy (SG), Roux-en-Y gastric bypass (RYGB), and biliopancreatic diversion with duodenal switch between 2016 and 2020 were selected from the Metabolic and Bariatric Surgery Accreditation and Quality Improvement Program (MBSAQIP) database. Patient variables reported by the MBSAQIP database were analyzed by ML algorithms random forest (RF), gradient boosting (XGB), and deep neural networks (NN), and the results of the predictive models were compared to logistic regression using area under the receiver operating characteristic curve (AUROC). RESULTS: Our study included 863,348 patients, of which 39,068 (4.52%) were readmitted. AUROC scores were XGB .785 (95% CI .784-.786), RF .785 (95% CI .784-.785), and NN .754 (95% CI .753-.754), compared with .62 (95% CI .62-.621) for logistic regression (LR) (P < .001). The sensitivity and specificity for XGB, the best performing model, were 73.81% and 70%, compared with 52.94% and 70% for logistic regression. The most important variables were intervention or reoperation prior to discharge, unplanned ICU admission, initial procedure, and the intraoperative transfusion. CONCLUSIONS: ML demonstrates significant advantages over logistic regression when predicting 30-day readmission following bariatric surgery. With external validation, models could identify the best candidates for early discharge or targeted postdischarge resources.

Application of machine learning to predict postoperative gastrointestinal bleed in bariatric surgery.

BACKGROUND: Postoperative gastrointestinal bleeding (GIB) is a rare but serious complication of bariatric surgery. The recent rise in extended venous thromboembolism regimens as well as outpatient bariatric surgery may increase the risk of postoperative GIB or lead to delay in diagnosis. This study seeks to use machine learning (ML) to create a model that predicts postoperative GIB to aid surgeon decision-making and improve patient counseling for postoperative bleeds. METHODS: The Metabolic and Bariatric Surgery Accreditation and Quality Improvement Program (MBSAQIP) database was used to train and validate three types of ML methods: random forest (RF), gradient boosting (XGB), and deep neural networks (NN), and compare them with logistic regression (LR) regarding postoperative GIB. The dataset was split using fivefold cross-validation into training and validation sets, in an 80/20 ratio. The performance of the models was assessed using area under the receiver operating characteristic curve (AUROC) and compared with the DeLong test. Variables with the strongest effect were identified using Shapley additive explanations (SHAP). RESULTS: The study included 159,959 patients. Postoperative GIB was identified in 632 (0.4%) patients. The three ML methods, RF (AUROC 0.764), XGB (AUROC 0.746), and NN (AUROC 0.741) all outperformed LR (AUROC 0.709). The best ML method, RF, was able to predict postoperative GIB with a specificity and sensitivity of 70.0% and 75.4%, respectively. Using DeLong testing, the difference between RF and LR was determined to be significant with p < 0.01. Type of bariatric surgery, pre-op hematocrit, age, duration of procedure, and pre-op creatinine were the 5 most important features identified by ML retrospectively. CONCLUSIONS: We have developed a ML model that outperformed LR in predicting postoperative GIB. Using ML models for risk prediction can be a helpful tool for both surgeons and patients undergoing bariatric procedures but more interpretable models are needed.

Prediction of Ureteral Injury During Colorectal Surgery Using Machine Learning.

BACKGROUND: Ureteral injury (UI) is a rare but devastating complication during colorectal surgery. Ureteral stents may reduce UI but carry risks themselves. Risk predictors for UI could help target the use of stents, but previous efforts have relied on logistic regression (LR), shown moderate accuracy, and used intraoperative variables. We sought to use an emerging approach in predictive analytics, machine learning, to create a model for UI. METHODS: Patients who underwent colorectal surgery were identified in the National Surgical Quality Improvement Program (NSQIP) database. Patients were split into training, validation, and test sets. The primary outcome was UI. Three machine learning approaches were tested including random forest (RF), gradient boosting (XGB), and neural networks (NN), and compared with traditional LR. Model performance was assessed using area under the curve (AUROC). RESULTS: The data set included 262,923 patients, of whom 1519 (.578%) experienced UI. Of the modeling techniques, XGB performed the best, with an AUROC score of .774 (95% CI .742-.807) compared with .698 (95% CI .664-.733) for LR. Random forest and NN performed similarly with scores of .738 and .763, respectively. Type of procedure, work RVUs, indication for surgery, and mechanical bowel prep showed the strongest influence on model predictions. CONCLUSIONS: Machine learning-based models significantly outperformed LR and previous models and showed high accuracy in predicting UI during colorectal surgery. With proper validation, they could be used to support decision making regarding the placement of ureteral stents preoperatively.

Computer Vision Analysis of Specimen Mammography to Predict Margin Status.

Intra-operative specimen mammography is a valuable tool in breast cancer surgery, providing immediate assessment of margins for a resected tumor. However, the accuracy of specimen mammography in detecting microscopic margin positivity is low. We sought to develop a deep learning-based model to predict the pathologic margin status of resected breast tumors using specimen mammography. A dataset of specimen mammography images matched with pathology reports describing margin status was collected. Models pre-trained on radiologic images were developed and compared with models pre-trained on non-medical images. Model performance was assessed using sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC). The dataset included 821 images and 53% had positive margins. For three out of four model architectures tested, models pre-trained on radiologic images outperformed domain-agnostic models. The highest performing model, InceptionV3, showed a sensitivity of 84%, a specificity of 42%, and AUROC of 0.71. These results compare favorably with the published literature on surgeon and radiologist interpretation of specimen mammography. With further development, these models could assist clinicians with identifying positive margins intra-operatively and decrease the rate of positive margins and re-operation in breast-conserving surgery.

Improved Prediction of Surgical-Site Infection After Colorectal Surgery Using Machine Learning.

BACKGROUND: Surgical-site infection is a source of significant morbidity after colorectal surgery. Previous efforts to develop models that predict surgical-site infection have had limited accuracy. Machine learning has shown promise in predicting postoperative outcomes by identifying nonlinear patterns within large data sets. OBJECTIVE: This study aimed to seek usage of machine learning to develop a more accurate predictive model for colorectal surgical-site infections. DESIGN: Patients who underwent colorectal surgery were identified in the American College of Surgeons National Quality Improvement Program database from years 2012 to 2019 and were split into training, validation, and test sets. Machine-learning techniques included random forest, gradient boosting, and artificial neural network. A logistic regression model was also created. Model performance was assessed using area under the receiver operating characteristic curve. SETTINGS: A national, multicenter data set. PATIENTS: Patients who underwent colorectal surgery. MAIN OUTCOME MEASURES: The primary outcome (surgical-site infection) included patients who experienced superficial, deep, or organ-space surgical-site infections. RESULTS: The data set included 275,152 patients after the application of exclusion criteria. Of all patients, 10.7% experienced a surgical-site infection. Artificial neural network showed the best performance with area under the receiver operating characteristic curve of 0.769 (95% CI, 0.762-0.777), compared with 0.766 (95% CI, 0.759-0.774) for gradient boosting, 0.764 (95% CI, 0.756-0.772) for random forest, and 0.677 (95% CI, 0.669-0.685) for logistic regression. For the artificial neural network model, the strongest predictors of surgical-site infection were organ-space surgical-site infection present at time of surgery, operative time, oral antibiotic bowel preparation, and surgical approach. LIMITATIONS: Local institutional validation was not performed. CONCLUSIONS: Machine-learning techniques predict colorectal surgical-site infections with higher accuracy than logistic regression. These techniques may be used to identify patients at increased risk and to target preventive interventions for surgical-site infection. See Video Abstract at http://links.lww.com/DCR/C88 . PREDICCIN MEJORADA DE LA INFECCIN DEL SITIO QUIRRGICO DESPUS DE LA CIRUGA COLORRECTAL MEDIANTE EL APRENDIZAJE AUTOMTICO: ANTECEDENTES:La infección del sitio quirúrgico es una fuente de morbilidad significativa después de la cirugía colorrectal. Los esfuerzos anteriores para desarrollar modelos que predijeran la infección del sitio quirúrgico han tenido una precisión limitada. El aprendizaje automático se ha mostrado prometedor en la predicción de los resultados posoperatorios mediante la identificación de patrones no lineales dentro de grandes conjuntos de datos.OBJETIVO:Intentamos utilizar el aprendizaje automático para desarrollar un modelo predictivo más preciso para las infecciones del sitio quirúrgico colorrectal.DISEÑO:Los pacientes que se sometieron a cirugía colorrectal se identificaron en la base de datos del Programa Nacional de Mejoramiento de la Calidad del Colegio Estadounidense de Cirujanos de los años 2012 a 2019 y se dividieron en conjuntos de capacitación, validación y prueba. Las técnicas de aprendizaje automático incluyeron conjunto aleatorio, aumento de gradiente y red neuronal artificial. También se creó un modelo de regresión logística. El rendimiento del modelo se evaluó utilizando el área bajo la curva característica operativa del receptor.CONFIGURACIÓN:Un conjunto de datos multicéntrico nacional.PACIENTES:Pacientes intervenidos de cirugía colorrectal.PRINCIPALES MEDIDAS DE RESULTADO:El resultado primario (infección del sitio quirúrgico) incluyó pacientes que experimentaron infecciones superficiales, profundas o del espacio de órganos del sitio quirúrgico.RESULTADOS:El conjunto de datos incluyó 275.152 pacientes después de la aplicación de los criterios de exclusión. El 10,7% de los pacientes presentó infección del sitio quirúrgico. La red neuronal artificial mostró el mejor rendimiento con el área bajo la curva característica operativa del receptor de 0,769 (IC del 95 %: 0,762 - 0,777), en comparación con 0,766 (IC del 95 %: 0,759 - 0,774) para el aumento de gradiente, 0,764 (IC del 95 %: 0,756 - 0,772) para conjunto aleatorio y 0,677 (IC 95% 0,669 - 0,685) para regresión logística. Para el modelo de red neuronal artificial, los predictores más fuertes de infección del sitio quirúrgico fueron la infección del sitio quirúrgico del espacio del órgano presente en el momento de la cirugía, el tiempo operatorio, la preparación intestinal con antibióticos orales y el abordaje quirúrgico.LIMITACIONES:No se realizó validación institucional local.CONCLUSIONES:Las técnicas de aprendizaje automático predicen infecciones del sitio quirúrgico colorrectal con mayor precisión que la regresión logística. Estas técnicas se pueden usar para identificar a los pacientes con mayor riesgo y para orientar las intervenciones preventivas para la infección del sitio quirúrgico. Consulte Video Resumen en http://links.lww.com/DCR/C88 . (Traducción-Dr Yolanda Colorado ).

Development and Validation of Machine Learning Models to Predict Readmission After Colorectal Surgery.

BACKGROUND: Readmission after colorectal surgery is common and often implies complications for patients and costs for hospitals. Previous works have created predictive models using logistic regression for this outcome but have shown limited accuracy. Machine learning has shown promise in improving predictions by identifying non-linear patterns in data. We sought to create a more accurate predictive model for readmission after colorectal surgery using machine learning. METHODS: Patients who underwent colorectal surgery were identified in the National Quality Improvement Program (NSQIP) database including years 2012-2019 and split into training, validation, and test sets. The primary outcome was readmission within 30 days of surgery. Three types of machine learning models were created, including random forest (RF), gradient boosting (XGB), and neural network (NN). A logistic regression (LR) model was also created for comparison. Model performance was evaluated using area under the receiver operating characteristic curve (AUROC). RESULTS: The dataset included 213,827 patients after application of exclusion criteria. A total of 23,083 (10.8%) of patients experienced readmission. NN obtained an AUROC of 0.751 (95% CI 0.743-0.759), compared with 0.684 (95% CI 0.676-0.693) for LR. RF and XGB performed similarly with AUROCs of 0.749 (95% CI 0.741-0.757) and 0.745 (95% CI 0.737-0.753) respectively. Ileus, index admission length of stay, organ-space surgical site infection present at time of surgery, and ostomy placement were identified as the most contributory variables. CONCLUSIONS: Machine learning approaches outperformed traditional statistical methods in the prediction of readmission after colorectal surgery. After external validation, this improved prediction model could be used to target interventions to reduce readmission rate.

Differential Performance of Machine Learning Models in Prediction of Procedure-Specific Outcomes.

BACKGROUND: Procedure-specific complications can have devastating consequences. Machine learning-based tools have the potential to outperform traditional statistical modeling in predicting their risk and guiding decision-making. We sought to develop and compare deep neural network (NN) models, a type of machine learning, to logistic regression (LR) for predicting anastomotic leak after colectomy, bile leak after hepatectomy, and pancreatic fistula after pancreaticoduodenectomy (PD). METHODS: The colectomy, hepatectomy, and PD National Surgical Quality Improvement Program (NSQIP) databases were analyzed. Each dataset was split into training, validation, and testing sets in a 60/20/20 ratio, with fivefold cross-validation. Models were created using NN and LR for each outcome. Models were evaluated primarily with area under the receiver operating characteristic curve (AUROC). RESULTS: A total of 197,488 patients were included for colectomy, 25,403 for hepatectomy, and 23,333 for PD. For anastomotic leak, AUROC for NN was 0.676 (95% 0.666-0.687), compared with 0.633 (95% CI 0.620-0.647) for LR. For bile leak, AUROC for NN was 0.750 (95% CI 0.739-0.761), compared with 0.722 (95% CI 0.698-0.746) for LR. For pancreatic fistula, AUROC for NN was 0.746 (95% CI 0.733-0.760), compared with 0.713 (95% CI 0.703-0.723) for LR. Variables related to intra-operative information, such as surgical approach, biliary reconstruction, and pancreatic gland texture were highly important for model predictions. DISCUSSION: Machine learning showed a marginal advantage over traditional statistical techniques in predicting procedure-specific outcomes. However, models that included intra-operative information performed better than those that did not, suggesting that NSQIP procedure-targeted datasets may be strengthened by including relevant intra-operative information.