Dataset Information

Probabilistic forecasting of surgical case duration using machine learning: model development and validation.

ABSTRACT:

Objective

Accurate estimations of surgical case durations can lead to the cost-effective utilization of operating rooms. We developed a novel machine learning approach, using both structured and unstructured features as input, to predict a continuous probability distribution of surgical case durations.

Materials and methods

The data set consisted of 53 783 surgical cases performed over 4 years at a tertiary-care pediatric hospital. Features extracted included categorical (American Society of Anesthesiologists [ASA] Physical Status, inpatient status, day of week), continuous (scheduled surgery duration, patient age), and unstructured text (procedure name, surgical diagnosis) variables. A mixture density network (MDN) was trained and compared to multiple tree-based methods and a Bayesian statistical method. A continuous ranked probability score (CRPS), a generalized extension of mean absolute error, was the primary performance measure. Pinball loss (PL) was calculated to assess accuracy at specific quantiles. Performance measures were additionally evaluated on common and rare surgical procedures. Permutation feature importance was measured for the best performing model.

Results

MDN had the best performance, with a CRPS of 18.1 minutes, compared to tree-based methods (19.5-22.1 minutes) and the Bayesian method (21.2 minutes). MDN had the best PL at all quantiles, and the best CRPS and PL for both common and rare procedures. Scheduled duration and procedure name were the most important features in the MDN.

Conclusions

Using natural language processing of surgical descriptors, we demonstrated the use of ML approaches to predict the continuous probability distribution of surgical case durations. The more discerning forecast of the ML-based MDN approach affords opportunities for guiding intelligent schedule design and day-of-surgery operational decisions.

SUBMITTER: Jiao Y

PROVIDER: S-EPMC7727362 | biostudies-literature |

REPOSITORIES: biostudies-literature

ACCESS DATA

Similar Datasets

Project description:Importance:Inpatient overcrowding is associated with delays in care, including the deferral of surgical care until beds are available to accommodate postoperative patients. Timely patient discharge is critical to address inpatient overcrowding and requires coordination among surgeons, nurses, case managers, and others. This is difficult to achieve without early identification and systemwide transparency of discharge candidates and their respective barriers to discharge. Objective:To validate the performance of a clinically interpretable feedforward neural network model that could improve the discharge process by predicting which patients would be discharged within 24 hours and their clinical and nonclinical barriers. Design, Setting, and Participants:This prognostic study included adult patients discharged from inpatient surgical care from May 1, 2016, to August 31, 2017, at a quaternary care teaching hospital. Model performance was assessed with standard cross-validation techniques. The model's performance was compared with a baseline model using historical procedure median length of stay to predict discharges. In prospective cohort analysis, the feedforward neural network model was used to make predictions on general surgical care floors with 63 beds. If patients were not discharged when predicted, the causes of delay were recorded. Main Outcomes and Measures:The primary outcome was the out-of-sample area under the receiver operating characteristic curve of the model. Secondary outcomes included the causes of discharge delay and the number of avoidable bed-days. Results:The model was trained on 15 201 patients (median [interquartile range] age, 60 [46-70] years; 7623 [50.1%] men) discharged from inpatient surgical care. The estimated out-of-sample area under the receiver operating characteristic curve of the model was 0.840 (SD, 0.008; 95% CI, 0.839-0.844). Compared with the baseline model, the neural network model had higher sensitivity (52.5% vs 56.6%) and specificity (51.7% vs 82.6%). The neural network model identified 65 barriers to discharge. In the prospective study of 605 patients, causes of delays included clinical barriers (41 patients [30.1%]), variation in clinical practice (30 patients [22.1%]), and nonclinical reasons (65 patients [47.8%]). Summing patients who were not discharged owing to variation in clinical practice and nonclinical reasons, 128 bed-days, or 1.2 beds per day, were classified as avoidable. Conclusions and Relevance:This cohort study found that a neural network model could predict daily inpatient surgical care discharges and their barriers. The model identified systemic causes of discharge delays. Such models should be studied for their ability to increase the timeliness of discharges.

Project description:BackgroundChangeful seasonal influenza activity in subtropical areas such as Taiwan causes problems in epidemic preparedness. The Taiwan Centers for Disease Control has maintained real-time national influenza surveillance systems since 2004. Except for timely monitoring, epidemic forecasting using the national influenza surveillance data can provide pivotal information for public health response.ObjectiveWe aimed to develop predictive models using machine learning to provide real-time influenza-like illness forecasts.MethodsUsing surveillance data of influenza-like illness visits from emergency departments (from the Real-Time Outbreak and Disease Surveillance System), outpatient departments (from the National Health Insurance database), and the records of patients with severe influenza with complications (from the National Notifiable Disease Surveillance System), we developed 4 machine learning models (autoregressive integrated moving average, random forest, support vector regression, and extreme gradient boosting) to produce weekly influenza-like illness predictions for a given week and 3 subsequent weeks. We established a framework of the machine learning models and used an ensemble approach called stacking to integrate these predictions. We trained the models using historical data from 2008-2014. We evaluated their predictive ability during 2015-2017 for each of the 4-week time periods using Pearson correlation, mean absolute percentage error (MAPE), and hit rate of trend prediction. A dashboard website was built to visualize the forecasts, and the results of real-world implementation of this forecasting framework in 2018 were evaluated using the same metrics.ResultsAll models could accurately predict the timing and magnitudes of the seasonal peaks in the then-current week (nowcast) (ρ=0.802-0.965; MAPE: 5.2%-9.2%; hit rate: 0.577-0.756), 1-week (ρ=0.803-0.918; MAPE: 8.3%-11.8%; hit rate: 0.643-0.747), 2-week (ρ=0.783-0.867; MAPE: 10.1%-15.3%; hit rate: 0.669-0.734), and 3-week forecasts (ρ=0.676-0.801; MAPE: 12.0%-18.9%; hit rate: 0.643-0.786), especially the ensemble model. In real-world implementation in 2018, the forecasting performance was still accurate in nowcasts (ρ=0.875-0.969; MAPE: 5.3%-8.0%; hit rate: 0.582-0.782) and remained satisfactory in 3-week forecasts (ρ=0.721-0.908; MAPE: 7.6%-13.5%; hit rate: 0.596-0.904).ConclusionsThis machine learning and ensemble approach can make accurate, real-time influenza-like illness forecasts for a 4-week period, and thus, facilitate decision making.

Project description:BackgroundThe recent COVID-19 pandemic highlighted the challenges for traditional forecasting. Prediction markets are a promising way to generate collective forecasts and could potentially be enhanced if high-quality crowdsourced inputs were identified and preferentially weighted for likely accuracy in real-time with machine learning.MethodsWe aim to leverage human prediction markets with real-time machine weighting of likely higher accuracy trades to improve performance. The crowd sourced Almanis prediction market longitudinal platform (n = 1822) and Next Generation Social Science (NGS2) platform (n = 103) were utilised.FindingsA 43-feature model predicted accurate forecasters, those with top quintile relative Brier accuracy, with subsequent replication in two out-of-sample datasets (pboth <1 × 10-9). Trades graded by this model as having higher accuracy scores than others produced a greater AUC temporal gain in the overall market after vs before trade. Accuracy score-weighted forecasts had higher accuracy than market forecasts alone, particularly when the two systems disagreed by 5% or more for binary event prediction: the hybrid system demonstrating substantial % AUC gains of 13.2%, p = 1.35 × 10-14 and 13.8%, p = 0.003 in two out-of-sample datasets. When discordant, the hybrid model was correct for COVID-19 event occurrence 72.7% of the time vs 27.3% for market models, p = 0.007. This net classification benefit was replicated in the separate Almanis B dataset, p = 2.4 × 10-7.InterpretationReal-time machine classification followed by weighting human trades according to likely accuracy improves collective forecasting performance. This could provide improved anticipation of and thus response to emerging risks.FundingThis work was supported by an AusIndustry R and D tax incentive program from the Department of Industry, Science, Energy and Resources, Australia, to SlowVoice Pty Ltd. (IR 2101990) and Fellowship (GNT 1110200) and Investigator grant (GNT 1197234) to A-L Ponsonby by the National Health and Medical Research Council of Australia.

Project description:Background:Screening for Barrett's Oesophagus (BE) relies on endoscopy which is invasive and has a low yield. This study aimed to develop and externally validate a simple symptom and risk-factor questionnaire to screen for patients with BE. Methods:Questionnaires from 1299 patients in the BEST2 case-controlled study were analysed: 880 had BE including 40 with invasive oesophageal adenocarcinoma (OAC) and 419 were controls. This was randomly split into a training cohort of 776 patients and an internal validation cohort of 523 patients. External validation included 398 patients from the BOOST case-controlled study: 198 with BE (23 with OAC) and 200 controls. Identification of independently important diagnostic features was undertaken using machine learning techniques information gain (IG) and correlation based feature selection (CFS). Multiple classification tools were assessed to create a multi-variable risk prediction model. Internal validation was followed by external validation in the independent dataset. Findings:The BEST2 study included 40 features. Of these, 24 added IG but following CFS, only 8 demonstrated independent diagnostic value including age, gender, smoking, waist circumference, frequency of stomach pain, duration of heartburn and acid taste and taking of acid suppression medicines. Logistic regression offered the highest prediction quality with AUC (area under the receiver operator curve) of 0.87. In the internal validation set, AUC was 0.86. In the BOOST external validation set, AUC was 0.81. Interpretation:The diagnostic model offers valid predictions of diagnosis of BE in patients with symptomatic gastroesophageal reflux, assisting in identifying who should go forward to invasive testing. Overweight men who have been taking stomach medicines for a long time may merit particular consideration for further testing. The risk prediction tool is quick and simple to administer but will need further calibration and validation in a prospective study in primary care. Funding:Charles Wolfson Trust and Guts UK.