Artificial Intelligence in Surgical Research: Accomplishments and Future Directions.

MINI-ABSTRACT: The study introduces various methods of performing conventional ML and their implementation in surgical areas, and the need to move beyond these traditional approaches given the advent of big data. OBJECTIVE: Investigate current understanding and future directions of machine learning applications, such as risk stratification, clinical data analytics, and decision support, in surgical practice. SUMMARY BACKGROUND DATA: The advent of the electronic health record, near unlimited computing, and open-source computational packages have created an environment for applying artificial intelligence, machine learning, and predictive analytic techniques to healthcare. The "hype" phase has passed, and algorithmic approaches are being developed for surgery patients through all stages of care, involving preoperative, intraoperative, and postoperative components. Surgeons must understand and critically evaluate the strengths and weaknesses of these methodologies. METHODS: The current body of AI literature was reviewed, emphasizing on contemporary approaches important in the surgical realm. RESULTS AND CONCLUSIONS: The unrealized impacts of AI on clinical surgery and its subspecialties are immense. As this technology continues to pervade surgical literature and clinical applications, knowledge of its inner workings and shortcomings is paramount in determining its appropriate implementation.

A neural network analysis of the effect of high and low frailty index indicators on predicting elective surgery discharge destinations.

BACKGROUND: Frailty is frequently used by clinicians to help determine surgical outcomes. The frailty index, which represents the frequency of frailty indicators present in an individual, is one method for evaluating patient frailty to predict surgical outcomes. However, the frailty index treats all indicators of frailty that are used in the index as equivalent. Our hypothesis is that frailty indicators may be divided into groups of high and low-impact indicators and this separation will improve surgical discharge outcome prediction accuracy. DATA AND METHODS: Population data for inpatient elective operations was collected from the 2018 American College of Surgeons National Surgical Quality Improvement Program Participant Use Files. Artificial neural network (ANN) models trained using backpropagation are used to evaluate the relative accuracy for predicting surgical outcome of discharge destination using a traditional modified frailty index (mFI) or a new joint mFI that separates high-impact and low-impact indicators into distinct groups as input variables. Predictions are made across nine possible discharge destinations. A leave-one-out method is used to indicate the relative contribution of high and low-impact variables. RESULTS: Except for the surgical specialty of cardiac surgery, the ANN model using distinct high and low-impact mFI indexes uniformly outperformed the ANN models using a single traditional mFI. Prediction accuracy improved from 3.4% to 28.1%. The leave-one-out experiment shows that except for the case of otolaryngology operations, the high-impact index indicators provided more support when determining surgical discharge destination outcomes. CONCLUSION: Frailty indicators are not uniformly similar and should be treated differently in clinical outcome prediction systems.

Predicting Elective Surgical Patient Outcome Destination Based on the Preoperative Modified Frailty Index and Laboratory Values.

INTRODUCTION: To determine the accuracy of preoperative modified frailty index (mFI) with or without laboratory values (mFI-labs or labs-continuous) in predicting postoperative discharge destination. Discharge destination is important to providers and patients. The ability to accurately predict discharge destination preoperatively can improve hospital resource utilization and help set patient and family expectations. METHODS: Cohort analysis of the 2018 American College of Surgeon National Surgical Quality Improvement Project (ACS-NSQIP) Participant Use File of patients undergoing operations with complete data point sets: age, sex, operation work relative-value units; mFI-clinical based on 12 clinical findings, mFI-labs based on seven laboratory values. The nine hierarchical destinations: home, home with assistance, multi-level community, unskilled-care facility, rehabilitation facility, skilled-nursing facility, acute care hospital, hospice, or death, from best to worst outcome. Data were analyzed using univariate analysis, multiple logistic regression and supervised learning artificial neural networks. RESULTS: Univariate and multivariate in general showed that patients with higher mFI-clinical and mFI-lab scores, as well as older age and more complex operations were more likely to be discharged to facilities other than home. However, these statistical techniques could not predict the exact destination. An artificial neural network analysis demonstrated perfect location prediction in 64.9% of cases and within one level of prefect prediction is 87.4%. CONCLUSIONS: Using a limited number of preoperative factors, combining the mFI-clinical with laboratory values significantly improves the destination prediction performance significantly better than using the values separately. Preoperative knowledge of the likely discharge destination can benefit postoperative care planning and delivery.

Predicting Crime and Other Uses of Neural Networks in Police Decision Making.

Neural networks are a machine learning method that excel in solving classification and forecasting problems. They have also been shown to be a useful tool for working with big data oriented environments such as law enforcement. This article reviews and examines existing research on the utilization of neural networks for forecasting crime and other police decision making problem solving. Neural network models to predict specific types of crime using location and time information and to predict a crime's location when given the crime and time of day are developed to demonstrate the application of neural networks to police decision making. The neural network crime prediction models utilize geo-spatiality to provide immediate information on crimes to enhance law enforcement decision making. The neural network models are able to predict the type of crime being committed 16.4% of the time for 27 different types of crime or 27.1% of the time when similar crimes are grouped into seven categories of crime. The location prediction neural networks are able to predict the zip code location or adjacent location 31.2% of the time.

Prediction of perioperative transfusions using an artificial neural network.

BACKGROUND: Accurate prediction of operative transfusions is essential for resource allocation and identifying patients at risk of postoperative adverse events. This research examines the efficacy of using artificial neural networks (ANNs) to predict transfusions for all inpatient operations. METHODS: Over 1.6 million surgical cases over a two year period from the NSQIP-PUF database are used. Data from 2014 (750937 records) are used for model development and data from 2015 (885502 records) are used for model validation. ANN and regression models are developed to predict perioperative transfusions for surgical patients. RESULTS: Various ANN models and logistic regression, using four variable sets, are compared. The best performing ANN models with respect to both sensitivity and area under the receiver operator characteristic curve outperformed all of the regression models (p < .001) and achieved a performance of 70-80% specificity with a corresponding 75-62% sensitivity. CONCLUSION: ANNs can predict >75% of the patients who will require transfusion and 70% of those who will not. Increasing specificity to 80% still enables a sensitivity of almost 67%. The unique contribution of this research is the utilization of a single ANN model to predict transfusions across a broad range of surgical procedures.

Cyber Attacks on Healthcare Devices Using Unmanned Aerial Vehicles.

The growing use of wireless technology in healthcare systems and devices makes these systems particularly open to cyber-based attacks, including denial of service and information theft via sniffing (eaves-dropping) and phishing attacks. Evolving technology enables wireless healthcare systems to communicate over longer ranges, which opens them up to greater numbers of possible threats. Unmanned aerial vehicles (UAV) or drones present a new and evolving attack surface for compromising wireless healthcare systems. An enumeration of the types of wireless attacks capable via drones are presented, including two new types of cyber threats: a stepping stone attack and a cloud-enabled attack. A real UAV is developed to test and demonstrate the vulnerabilities of healthcare systems to this new threat vector. The UAV successfully attacked a simulated smart hospital environment and also a small collection of wearable healthcare sensors. Compromise of wearable or implanted medical devices can lead to increased morbidity and mortality.

Prophylactic antibiotic bundle compliance and surgical site infections: an artificial neural network analysis.

BACKGROUND: Best practice "bundles" have been developed to lower the occurrence rate of surgical site infections (SSI's). We developed artificial neural network (ANN) models to predict SSI occurrence based on prophylactic antibiotic compliance. METHODS: Using the American College of Surgeons National Quality Improvement Program (ACS-NSQIP) Tampa General Hospital patient dataset for a six-month period, 780 surgical procedures were reviewed for compliance with SSI guidelines for antibiotic type and timing. SSI rates were determined for patients in the compliant and non-compliant groups. ANN training and validation models were developed to include the variables of age, sex, steroid use, bleeding disorders, transfusion, white blood cell count, hematocrit level, platelet count, wound class, ASA class, and surgical antimicrobial prophylaxis (SAP) bundle compliance. RESULTS: Overall compliance to recommended antibiotic type and timing was 92.0%. Antibiotic bundle compliance had a lower incidence of SSI's (3.3%) compared to the non-compliant group (8.1%, p = 0.07). ANN models predicted SSI with a 69-90% sensitivity and 50-60% specificity. The model was more sensitive when bundle compliance was not used in the model, but more specific when it was. Preoperative white blood cell (WBC) count had the most influence on the model. CONCLUSIONS: SAP bundle compliance was associated with a lower incidence of SSI's. In an ANN model, inclusion of the SAP bundle compliance reduced sensitivity, but increased specificity of the prediction model. Preoperative WBC count had the most influence on the model.

An Evaluation of Artificial Neural Networks in Predicting Pancreatic Cancer Survival.

OBJECTIVE: This study aims to evaluate the development of an artificial neural network (ANN) method for predicting the survival likelihood of pancreatic adenocarcinoma patients. The ANN predictive model should produce results with a 90% sensitivity. METHODS: A prospective examination of the records for 283 consecutive pancreatic adenocarcinoma patients is used to identify 219 records with complete data. These records are then used to create two unique samples which are then used to train and validate an ANN predictive model. Numerous network architectures are evaluated, following recommended ANN development protocols. RESULTS: Several backpropagation-trained ANNs were produced that satisfied the 90% sensitivity requirement. An ANN model with over a 91% sensitivity is selected because even though it did not have the highest sensitivity, it was able to achieve over 38% specificity. CONCLUSIONS: ANN models can accurately predict the 7-month survival of pancreatic adenocarcinoma patients, both with and without resection, at a 91% sensitivity and 38% specificity. This implies that ANN models may be useful objective decision tools in complex treatment decisions. This information may be used by patients and surgeons in determining optimal treatment plans that minimize regret and improve the quality of life for these patients.

Impacts of mobile tablet computing on provider productivity, communications, and the process of care.

OBJECTIVE: Health information technology investments continue to increase while the value derived from their implementation and use is mixed. Mobile device adoption into practice is a recent trend that has increased dramatically and formal studies are needed to investigate consequent benefits and challenges. The objective of this study is to evaluate practitioner perceptions of improvements in productivity, provider-patient communications, care provision, technology usability and other outcomes following the adoption and use of a tablet computer connected to electronic health information resources. METHODS: A pilot program was initiated in June 2013 to evaluate the effect of mobile tablet computers at one health provider organization in the southeast United States. Providers were asked to volunteer for the evaluation and were each given a mobile tablet computer. A total of 42 inpatient and outpatient providers were interviewed in 2015 using a survey style questionnaire that utilized yes/no, Likert-style, and open ended questions. Each had previously used an electronic health record (EHR) system a minimum of one year outside of residency, and were regular users of personal mobile devices. Each used a mobile tablet computer in the context of their practice connected to the health system EHR. RESULTS: The survey results indicate that more than half of providers perceive the use of the tablet device as having a positive effect on patient communications, patient education, patient's perception of the provider, time spent interacting with patients, provider productivity, process of care, satisfaction with EHR when used together with the device, and care provision. Providers also reported feeling comfortable using the device (82.9%), would recommend the device to colleagues (69.2%), did not experience increased information security and privacy concerns (95%), and noted significant reductions in EHR login times (64.1%). Less than 25% of participants reported negative impacts on any of these areas as well as on time spent on order submission, note completion time, overall workload, patient satisfaction with care experience and patient outcomes. Gender, number of years in practice, practice type (general practitioner vs. specialist), and service type (inpatient/outpatient) were found to have a significant effect on perceptions of patient satisfaction, care process, and provider productivity. CONCLUSIONS: Providers found positive gains from utilizing mobile devices in overall productivity, improved communications with their patients, the process of care, and technology efficiencies when used in combination with EHR and other health information resources. Demographic and health care work environment play a role in how mobile technologies are integrated into practice by providers.

Historical return on investment and improved quality resulting from development and mining of a hospital laboratory relational database.

A hospital laboratory relational database, developed over eight years, has demonstrated significant cost savings and a substantial financial return on investment (ROI). In addition, the database has been used to measurably improve laboratory operations and the quality of patient care.

Artificial neural network medical decision support tool: predicting transfusion requirements of ER patients.

Blood product transfusion is a financial concern for hospitals and patients. Efficient utilization of this dwindling resource is a critical problem if hospitals are to maximize patient care while minimizing costs. Traditional statistical models do not perform well in this domain. An additional concern is the speed with which transfusion decisions and planning can be made. Rapid assessment in the emergency room (ER) necessarily limits the amount of usable information available (with respect to independent variables available). This study evaluates the efficacy of using artificial neural networks (ANNs) to predict the transfusion requirements of trauma patients using readily available information. A total of 1016 patient records are used to train and test a backpropagation neural network for predicting the transfusion requirements of these patients during the first 2, 2-6, and 6-24 h, and for total transfusions. Sensitivity and specificity analysis are used along with the mean absolute difference between blood units predicted and units transfused to demonstrate that ANNs can accurately predict most ER patient transfusion requirements, while only using information available at the time of entry into the ER.

A multiagent architecture for developing medical information retrieval agents.

Information that is available on the world wide web (WWW) is already more vast than can be comprehensibly studied by individuals and this quantity is increasing at a staggering pace. The quality of service delivered by physicians is dependent on the availability of current information. The agent paradigm offers a means for enabling physicians to filter information and retrieve only information that is relevant to current patient treatments. As with many specialized domains, agent-based information retrieval in medical domains must satisfy several domain-dependent constraints. A multiple agent architecture is developed and described in detail to efficiently provide agent-based information retrieval from the WWW and other explicit information resources. A simulation of the proposed multiple agent architecture shows a 97% decrease in information overload and an 85% increase in information relevancy over existing meta-search tools (with even larger gains over standard search engines).

Factors affecting transfusion of fresh frozen plasma, platelets, and red blood cells during elective coronary artery bypass graft surgery.

CONTEXT: The ability to predict the use of blood components during surgery will improve the blood bank's ability to provide efficient service. OBJECTIVE: Develop prediction models using preoperative risk factors to assess blood component usage during elective coronary artery bypass graft surgery (CABG). DESIGN: Eighty-three preoperative, multidimensional risk variables were evaluated for patients undergoing elective CABG-only surgery. MAIN OUTCOMES MEASURES: The study endpoints included transfusion of fresh frozen plasma (FFP), platelets, and red blood cells (RBC). Multivariate logistic regression models were built to assess the predictors related to each of these endpoints. SETTING: Department of Veterans Affairs (VA) health care system. PATIENTS: Records for 3034 patients undergoing elective CABG-only procedures; 1033 patients received a blood component transfusion during CABG. RESULTS: Previous heart surgery and decreased ejection fraction were significant predictors of transfusion for all blood components. Platelet count was predictive of platelet transfusion and FFP utilization. Baseline hemoglobin was a predictive factor for more than 2 units of RBC. Some significant hospital variation was noted beyond that predicted by patient risk factors alone. CONCLUSIONS: Prediction models based on preoperative variables may facilitate blood component management for patients undergoing elective CABG. Algorithms are available to predict transfusion resources to assist blood banks in improving responsiveness to clinical needs. Predictors for use of each blood component may be identified prior to elective CABG for VA patients.

Clinical correlate of EEG rhythmicity.

The widespread use of the routine EEG in clinical practice was a major development in the treatment of patients with ill-defined spells thought to be epileptic. Not every finding on the EEG is suggestive of epilepsy, and the EEG is subject to over-interpretation, which may lead to misdiagnosis and incorrect treatment. Although supplemented by other procedures, the EEG remains a cost-effective and noninvasive way to diagnose spells. To enhance further the diagnostic use of the EEG, it is important to determine how strongly patterns are correlated with clinical seizures. The authors studied one EEG pattern, lateralized bursts of theta, and found the rhythmicity of the pattern to be most strongly correlated with seizures.