Research |
Improvement of patient care and hospital resource allocation
|
My research utilizes mathematical concepts of machine learning to improve patient outcomes and reduce hospital costs in Critical Care Medicine. This includes the development of algorithms aimed at improving (a) compliance in the intensive care unit (ICU), (b) patient outcome, and (c) bedside care. For example, in collaboration with medical doctors from the pediatric ICU, we have developed a math-based patient respiratory index that functions as a good predictor of patient outcome upon extubation/discharge (future need for non-invasive ventilation, future need of re-intubation, no need of respiratory assistance). This index is built using the last 2 - 6 hours of a patient's’ vital signs and respiratory variables recorded by the mechanical ventilator. Future development of patient "stability" indices will enable us to track, in real-time, milestones of patient's path to recovery and/or health deterioration.
|
|
Related publications
|
Adding Continuous Vital-Sign Information to Static Clinical Data Improves Prediction of Length of Stay Following Intubation. A Data Driven Machine Learning Approach
Castiñeira D, Schlosser K, Geva A, Rahmani A, Fiore G, Walsh BK, Smallwood CD, Arnold JH, Santillana M. Respiratory Care. 2020;65 (9). Abstract
BACKGROUND: Bedside monitors in the ICU routinely measure and collect patients' physiologic data in real time to continuously asess the health status of patients who are critically ill. With the advent of increased computational power and the ability to store and rapidly process big data sets in recent years, these physiologic data show promise in identifying specific outcomes and/or events during patients' ICU hospitalization. METHODS: We introduced a methodology designed to automatically extract information from continuous-in-time vital sign data collected from bedside monitors to predict if a patient will experience a prolonged stay (length of stay) on mechanical ventilation, defined as >4 d, in a pediatric ICU. RESULTS: Continuous-in-time vital signs information and clinical history data were retrospectively collected for 284 ICU subjects from their first 24 h on mechanical ventilation from a medical-surgical pediatric ICU at Boston Children's Hospital. Multiple machine learning models were trained on multiple subsets of these subjects to predict the likelihood that each of these subjects would experience a long stay. We evaluated the predictive power of our models strictly on unseen hold-out validation sets of subjects. Our methodology achieved model performance of >83% (area under the curve) by using only vital sign information as input, and performances of 90% (area under the curve) by combining vital sign information with subjects' static clinical data readily available in electronic health records. We implemented this approach on 300 independently trained experiments with different choices of training and hold-out validation sets to ensure the consistency and robustness of our results in our study sample. The predictive power of our approach outperformed recent efforts that used deep learning to predict a similar task. CONCLUSIONS: Our proposed workflow may prove useful in the design of scalable approaches for real-time predictive systems in ICU environments, exploiting real-time vital sign information from bedside monitors. (ClinicalTrials.gov registration NCT02184208.) |
|
Avoidable serum potassium testing in the cardiac intensive care unit: development and testing of a machine learning model.
Patel B, Sperotto F, Molina M, Kimura S, Delgado M, Santillana M, Kheir JN. Pediatric Critical Care Medicine. 2020;22 (4). Abstract
Objectives: To create a machine-learning model identifying potentially avoidable blood draws for serum potassium among pediatric patients following cardiac surgery. Design: Retrospective cohort study. Setting: Tertiary-care center. Patients: All patients admitted to the cardiac ICU at Boston Children’s Hospital between January 2010 and December 2018 with a length of stay greater than or equal to 4 days and greater than or equal to two recorded serum potassium measurements. Interventions: None. Measurements and Main Results: We collected variables related to potassium homeostasis, including serum chemistry, hourly potassium intake, diuretics, and urine output. Using established machine-learning techniques, including random forest classifiers, and hyperparameter tuning, we created models predicting whether a patient’s potassium would be normal or abnormal based on the most recent potassium level, medications administered, urine output, and markers of renal function. We developed multiple models based on different age-categories and temporal proximity of the most recent potassium measurement. We assessed the predictive performance of the models using an independent test set. Of the 7,269 admissions (6,196 patients) included, serum potassium was measured on average of 1 (interquartile range, 0–1) time per day. Approximately 96% of patients received at least one dose of IV diuretic and 83% received a form of potassium supplementation. Our models predicted a normal potassium value with a median positive predictive value of 0.900. A median percentage of 2.1% measurements (mean 2.5%; interquartile range, 1.3–3.7%) was incorrectly predicted as normal when they were abnormal. A median percentage of 0.0% (interquartile range, 0.0–0.4%) critically low or high measurements was incorrectly predicted as normal. A median of 27.2% (interquartile range, 7.8–32.4%) of samples was correctly predicted to be normal and could have been potentially avoided. Conclusions: Machine-learning methods can be used to predict avoidable blood tests accurately for serum potassium in critically ill pediatric patients. A median of 27.2% of samples could have been saved, with decreased costs and risk of infection or anemia. |
|
Machine learning approaches to predicting no-shows in pediatric medical appointment
Dianbo Liu, Won-Yong Shin, Eli Sprecher, Kathleen Conroy, Omar Santiago, Gal Wachtel, Mauricio Santillana NPJ digital medicine 5 (1), 1-1112022 Abstract
Patients’ no-shows, scheduled but unattended medical appointments, have a direct negative impact on patients’ health, due to discontinuity of treatment and late presentation to care. They also lead to inefficient use of medical resources in hospitals and clinics. The ability to predict a likely no-show in advance could enable the design and implementation of interventions to reduce the risk of it happening, thus improving patients’ care and clinical resource allocation. In this study, we develop a new interpretable deep learning-based approach for predicting the risk of no-shows at the time when a medical appointment is first scheduled. The retrospective study was conducted in an academic pediatric teaching hospital with a 20% no-show rate. Our approach tackles several challenges in the design of a predictive model by (1) adopting a data imputation method for patients with missing information in their records (77% of the population), (2) exploiting local weather information to improve predictive accuracy, and (3) developing an interpretable approach that explains how a prediction is made for each individual patient. Our proposed neural network-based and logistic regression-based methods outperformed persistence baselines. In an unobserved set of patients, our method correctly identified 83% of no-shows at the time of scheduling and led to a false alert rate less than 17%. Our method is capable of producing meaningful predictions even when some information in a patient’s records is missing. We find that patients’ past no-show record is the strongest predictor. Finally, we discuss several potential interventions to reduce no-shows, such as scheduling appointments of high-risk patients at off-peak times, which can serve as starting point for further studies on no-show interventions. |
|
Noninvasive Ventilation Is Interrupted Frequently and Mostly Used at Night in the Pediatric Intensive Care Unit
Schlosser KR, Fiore GA, Smallwood CD, Griffin J, Geva A, Santillana M, Arnold JH. Respiratory Care. 2019;64 (9). Abstract
BACKGROUND: Noninvasive ventilation (NIV) is commonly used to support children with respiratory failure, but detailed patterns of real-world use are lacking. The aim of our study was to describe use patterns of NIV via electronic medical record (EMR) data. METHODS: We performed a retrospective electronic chart review in a tertiary care pediatric ICU in the United States. Subjects admitted to the pediatric ICU from 2014 to 2017 who were mechanically ventilated were included in the study. RESULTS: The median number of discrete device episodes, defined as a time on support without interruption, was 20 (interquartile range [IQR] 8–49) per subject. The median duration of bi-level positive airway pressure (BPAP) support prior to interruption was 6.3 h (IQR 2.4–10.4); the median duration of CPAP was 6 h (IQR 2.1–10.4). Interruptions to BPAP had a median duration of 6.3 h (IQR 2–15.5); interruptions to CPAP had a median duration of 8.6 h (IQR 2.2–16.8). Use of NIV followed a diurnal pattern, with 44% of BPAP and 42% of CPAP subjects initiating support between 7:00 PM and midnight, and 49% of BPAP and 46% of CPAP subjects stopping support between 5:00 AM and 10:00 AM CONCLUSIONS: NIV was frequently interrupted, and initiation and discontinuation of NIV follows a diurnal pattern. Use of EMR data collected for routine clinical care allowed the analysis of granular details of typical use patterns. Understanding NIV use patterns may be particularly important to understanding the burden of pediatric ICU bed utilization for nocturnal NIV. To our knowledge, this is the first study to examine in detail the use of pediatric NIV and to define diurnal use and frequent interruptions to support. |
|
Internet search query data improves forecasts of daily emergency department volume
Tideman S, Santillana M, Bickel J, Reis B. Journal of the American Medical Informatics Association. 2019;ocz154. Abstract
ObjectiveEmergency departments (EDs) are increasingly overcrowded. Forecasting patient visit volume is challenging. Reliable and accurate forecasting strategies may help improve resource allocation and mitigate the effects of overcrowding. Patterns related to weather, day of the week, season, and holidays have been previously used to forecast ED visits. Internet search activity has proven useful for predicting disease trends and offers a new opportunity to improve ED visit forecasting. This study tests whether Google search data and relevant statistical methods can improve the accuracy of ED volume forecasting compared with traditional data sources. Materials and MethodsSeven years of historical daily ED arrivals were collected from Boston Children’s Hospital. We used data from the public school calendar, National Oceanic and Atmospheric Administration, and Google Trends. Multiple linear models using LASSO (least absolute shrinkage and selection operator) for variable selection were created. The models were trained on 5 years of data and out-of-sample accuracy was judged using multiple error metrics on the final 2 years. ResultsAll data sources added complementary predictive power. Our baseline day-of-the-week model recorded average percent errors of 10.99%. Autoregressive terms, calendar and weather data reduced errors to 7.71%. Search volume data reduced errors to 7.58% theoretically preventing 4 improperly staffed days. DiscussionThe predictive power provided by the search volume data may stem from the ability to capture population-level interaction with events, such as winter storms and infectious diseases, that traditional data sources alone miss. ConclusionsThis study demonstrates that search volume data can meaningfully improve forecasting of ED visit volume and could help improve quality and reduce cost. |
