Machine learning based forecast for the prediction of inpatient bed demand.

BACKGROUND: Overcrowding is a serious problem that impacts the ability to provide optimal level of care in a timely manner. High patient volume is known to increase the boarding time at the emergency department (ED), as well as at post-anesthesia care unit (PACU). Furthermore, the same high volume increases inpatient bed transfer times, which causes delays in elective surgeries, increases the probability of near misses, patient safety incidents, and adverse events. OBJECTIVE: The purpose of this study is to develop a Machine Learning (ML) based strategy to predict weekly forecasts of the inpatient bed demand in order to assist the resource planning for the ED and PACU, resulting in a more efficient utilization. METHODS: The data utilized included all adult inpatient encounters at Geisinger Medical Center (GMC) for the last 5 years. The variables considered were class of inpatient encounter, observation, or surgical overnight recovery (SORU) at the time of their discharge. The ML based strategy is built using the K-means clustering method and the Support Vector Machine Regression technique (K-SVR). RESULTS: The performance obtained by the K-SVR strategy in the retrospective cohort amounts to a mean absolute percentage error (MAPE) that ranges between 0.49 and 4.10% based on the test period. Additionally, results present a reduced variability, which translates into more stable forecasting results. CONCLUSIONS: The results from this study demonstrate the capacity of ML techniques to forecast inpatient bed demand, particularly using K-SVR. It is expected that the implementation of this model in the workflow of bed capacity management will create efficiencies, which will translate in a more reliable, inexpensive and timely care for patients.

Comparison of Machine Learning Algorithms for the Prediction of Preventable Hospital Readmissions.

A diverse universe of statistical models in the literature aim to help hospitals understand the risk factors of their preventable readmissions. However, these models are usually not necessarily applicable in other contexts, fail to achieve good discriminatory power, or cannot be compared with other models. We built and compared predictive models based on machine learning algorithms for 30-day preventable hospital readmissions of Medicare patients. This work used the same inclusion/exclusion criteria for diseases used by the Centers for Medicare and Medicaid Services. In addition, risk stratification techniques were implemented to study covariate behavior on each risk strata. The new models resulted in improved performance measured by the area under the receiver operating characteristic curve. Finally, factors such as higher length of stay, disease severity index, being discharged to a hospital, and primary language other than English were associated with increased risk to be readmitted within 30 days. In the future, better predictive models for 30-day preventable hospital readmissions can point to the development of systems that identify patients at high risk and lead to the implementation of interventions (e.g., discharge planning and follow-up) to those patients, providing consistent improvement in the quality and efficiency of the healthcare system.

A user needs assessment to inform health information exchange design and implementation.

BACKGROUND: Important barriers for widespread use of health information exchange (HIE) are usability and interface issues. However, most HIEs are implemented without performing a needs assessment with the end users, healthcare providers. We performed a user needs assessment for the process of obtaining clinical information from other health care organizations about a hospitalized patient and identified the types of information most valued for medical decision-making. METHODS: Quantitative and qualitative analysis were used to evaluate the process to obtain and use outside clinical information (OI) using semi-structured interviews (16 internists), direct observation (750 h), and operational data from the electronic medical records (30,461 hospitalizations) of an internal medicine department in a public, teaching hospital in Tampa, Florida. RESULTS: 13.7 % of hospitalizations generate at least one request for OI. On average, the process comprised 13 steps, 6 decisions points, and 4 different participants. Physicians estimate that the average time to receive OI is 18 h. Physicians perceived that OI received is not useful 33-66 % of the time because information received is irrelevant or not timely. Technical barriers to OI use included poor accessibility and ineffective information visualization. Common problems with the process were receiving extraneous notes and the need to re-request the information. Drivers for OI use were to trend lab or imaging abnormalities, understand medical history of critically ill or hospital-to-hospital transferred patients, and assess previous echocardiograms and bacterial cultures. About 85 % of the physicians believe HIE would have a positive effect on improving healthcare delivery. CONCLUSIONS: Although hospitalists are challenged by a complex process to obtain OI, they recognize the value of specific information for enhancing medical decision-making. HIE systems are likely to have increased utilization and effectiveness if specific patient-level clinical information is delivered at the right time to the right users.