Objective: We sought to develop an intensive care unit research database applying automated techniques to aggregate high-resolution diagnostic and therapeutic data from a large, diverse population of adult intensive care unit patients. This freely available database is intended to support epidemiologic research in critical care medicine and serve as a resource to evaluate new clinical decision support and monitoring algorithms. Design: Data collection and retrospective analysis. Setting: All adult intensive care units (medical intensive care unit, surgical intensive care unit, cardiac care unit, cardiac surgery recovery unit) at a tertiary care hospital. Patients: Adult patients admitted to intensive care units between 2001 and 2007. Interventions: None. Measurements and Main Results: The Multiparameter Intelligent Monitoring in Intensive Care II (MIMIC-II) database consists of 25,328 intensive care unit stays. The investigators collected detailed information about intensive care unit patient stays, including laboratory data, therapeutic intervention profiles such as vasoactive medication drip rates and ventilator settings, nursing progress notes, discharge summaries, radiology reports, provider order entry data, International Classification of Diseases, 9th Revision codes, and, for a subset of patients, high-resolution vital sign trends and waveforms. Data were automatically deidentified to comply with Health Insurance Portability and Accountability Act standards and integrated with relational database software to create electronic intensive care unit records for each patient stay. The data were made freely available in February 2010 through the Internet along with a detailed user's guide and an assortment of data processing tools. The overall hospital mortality rate was 11.7%, which varied by critical care unit. The median intensive care unit length of stay was 2.2 days (interquartile range, 1.1–4.4 days). According to the primary International Classification of Diseases, 9th Revision codes, the following disease categories each comprised at least 5% of the case records: diseases of the circulatory system (39.1%); trauma (10.2%); diseases of the digestive system (9.7%); pulmonary diseases (9.0%); infectious diseases (7.0%); and neoplasms (6.8%). Conclusions: MIMIC-II documents a diverse and very large population of intensive care unit patient stays and contains comprehensive and detailed clinical data, including physiological waveforms and minute-by-minute trends for a subset of records. It establishes a new public-access resource for critical care research, supporting a diverse range of analytic studies spanning epidemiology, clinical decision-rule development, and electronic tool development.

Background: Text-based patient medical records are a vital resource in medical research. In order to preserve patient confidentiality, however, the U.S. Health Insurance Portability and Accountability Act (HIPAA) requires that protected health information (PHI) be removed from medical records before they can be disseminated. Manual de-identification of large medical record databases is prohibitively expensive, time-consuming and prone to error, necessitating automatic methods for large-scale, automated de-identification. Methods: We describe an automated Perl-based de-identification software package that is generally usable on most free-text medical records, e.g., nursing notes, discharge summaries, X-ray reports, etc. The software uses lexical look-up tables, regular expressions, and simple heuristics to locate both HIPAA PHI, and an extended PHI set that includes doctors' names and years of dates. To develop the de-identification approach, we assembled a gold standard corpus of re-identified nursing notes with real PHI replaced by realistic surrogate information. This corpus consists of 2,434 nursing notes containing 334,000 words and a total of 1,779 instances of PHI taken from 163 randomly selected patient records. This gold standard corpus was used to refine the algorithm and measure its sensitivity. To test the algorithm on data not used in its development, we constructed a second test corpus of 1,836 nursing notes containing 296,400 words. The algorithm's false negative rate was evaluated using this test corpus. Results: Performance evaluation of the de-identification software on the development corpus yielded an overall recall of 0.967, precision value of 0.749, and fallout value of approximately 0.002. On the test corpus, a total of 90 instances of false negatives were found, or 27 per 100,000 word count, with an estimated recall of 0.943. Only one full date and one age over 89 were missed. No patient names were missed in either corpus. Conclusion: We have developed a pattern-matching de-identification system based on dictionary look-ups, regular expressions, and heuristics. Evaluation based on two different sets of nursing notes collected from a U.S. hospital suggests that, in terms of recall, the software out-performs a single human de-identifier (0.81) and performs at least as well as a consensus of two human de-identifiers (0.94). The system is currently tuned to de-identify PHI in nursing notes and discharge summaries but is sufficiently generalized and can be customized to handle text files of any format. Although the accuracy of the algorithm is high, it is probably insufficient to be used to publicly disseminate medical data. The open-source de-identification software and the gold standard re-identified corpus of medical records have therefore been made available to researchers via the PhysioNet website to encourage improvements in the algorithm.

Cardiovascular variables such as heart rate (HR) and blood pressure (BP) are regulated by an underlying control system, and therefore, the time series of these vital signs exhibit rich dynamical patterns of interaction in response to external perturbations (e.g., drug administration), as well as pathological states (e.g., onset of sepsis and hypotension). A question of interest is whether “similar” dynamical patterns can be identified across a heterogeneous patient cohort, and be used for prognosis of patients’ health and progress. In this paper, we used a switching vector autoregressive framework to systematically learn and identify a collection of vital sign time series dynamics, which are possibly recurrent within the same patient and may be shared across the entire cohort. We show that these dynamical behaviors can be used to characterize the physiological “state” of a patient. We validate our technique using simulated time series of the cardiovascular system, and human recordings of HR and BP time series from an orthostatic stress study with known postural states. Using the HR and BP dynamics of an intensive care unit (ICU) cohort of over 450 patients from the MIMIC II database, we demonstrate that the discovered cardiovascular dynamics are significantly associated with hospital mortality (dynamic modes 3 and 9, $p=0.001$ , $p=0.006$ from logistic regression after adjusting for the APACHE scores). Combining the dynamics of BP time series and SAPS-I or APACHE-III provided a more accurate assessment of patient survival/mortality in the hospital than using SAPS-I and APACHE-III alone ( $p=0.005$ and $p=0.045$ ). Our results suggest that the discovered dynamics of vital sign time series may contain additional prognostic value beyond that of the baseline acuity measures, and can potentially be used as an independent predictor of outcomes in the ICU.

The critical state of intensive care unit (ICU) patients demands close monitoring, and as a result a large volume of multi-parameter data is collected continuously. This represents a unique opportunity for researchers interested in clinical data mining. We sought to foster a more transparent and efficient intensive care research community by building a publicly available ICU database, namely Multiparameter Intelligent Monitoring in Intensive Care II (MIMIC-II). The data harnessed in MIMIC-II were collected from the ICUs of Beth Israel Deaconess Medical Center from 2001 to 2008 and represent 26,870 adult hospital admissions (version 2.6). MIMIC-II consists of two major components: clinical data and physiological waveforms. The clinical data, which include patient demographics, intravenous medication drip rates, and laboratory test results, were organized into a relational database. The physiological waveforms, including 125 Hz signals recorded at bedside and corresponding vital signs, were stored in an open-source format. MIMIC-II data were also deidentified in order to remove protected health information. Any interested researcher can gain access to MIMIC-II free of charge after signing a data use agreement and completing human subjects training. MIMIC-II can support a wide variety of research studies, ranging from the development of clinical decision support algorithms to retrospective clinical studies. We anticipate that MIMIC-II will be an invaluable resource for intensive care research by stimulating fair comparisons among different studies.

Background: Within the intensive care unit (ICU), arterial blood pressure (ABP) is typically recorded at different (and sometimes uneven) sampling frequencies, and from different sensors, and is often corrupted by different artifacts and noise which are often non-Gaussian, nonlinear and nonstationary. Extracting robust parameters from such signals, and providing confidences in the estimates is therefore difficult and requires an adaptive filtering approach which accounts for artifact types. Methods: Using a large ICU database, and over 6000 hours of simultaneously acquired electrocardiogram (ECG) and ABP waveforms sampled at 125 Hz from a 437 patient subset, we documented six general types of ABP artifact. We describe a new ABP signal quality index (SQI), based upon the combination of two previously reported signal quality measures weighted together. One index measures morphological normality, and the other degradation due to noise. After extracting a 6084-hour subset of clean data using our SQI, we evaluated a new robust tracking algorithm for estimating blood pressure and heart rate (HR) based upon a Kalman Filter (KF) with an update sequence modified by the KF innovation sequence and the value of the SQI. In order to do this, we have created six novel models of different categories of artifacts that we have identified in our ABP waveform data. These artifact models were then injected into clean ABP waveforms in a controlled manner. Clinical blood pressure (systolic, mean and diastolic) estimates were then made from the ABP waveforms for both clean and corrupted data. The mean absolute error for systolic, mean and diastolic blood pressure was then calculated for different levels of artifact pollution to provide estimates of expected errors given a single value of the SQI. Results: Our artifact models demonstrate that artifact types have differing effects on systolic, diastolic and mean ABP estimates. We show that, for most artifact types, diastolic ABP estimates are less noise-sensitive than mean ABP estimates, which in turn are more robust than systolic ABP estimates. We also show that our SQI can provide error bounds for both HR and ABP estimates. Conclusion: The KF/SQI-fusion method described in this article was shown to provide an accurate estimate of blood pressure and HR derived from the ABP waveform even in the presence of high levels of persistent noise and artifact, and during extreme bradycardia and tachycardia. Differences in error between artifact types, measurement sensors and the quality of the source signal can be factored into physiological estimation using an unbiased adaptive filter, signal innovation and signal quality measures.

We hypothesize that local customized modeling will provide more accurate mortality prediction than the current standard approach using existing scoring systems. Mortality prediction models were developed for two subsets of patients in Multi-parameter Intelligent Monitoring for Intensive Care (MIMIC), a public de-identified ICU database, and for the subset of patients >80 years old in a cardiac surgical patient registry. Logistic regression (LR), Bayesian network (BN) and artificial neural network (ANN) were employed. The best-fitted models were tested on the remaining unseen data and compared to either the Simplified Acute Physiology Score (SAPS) for the ICU patients, or the EuroSCORE for the cardiac surgery patients. Local customized mortality prediction models performed better as compared to the corresponding current standard severity scoring system for all three subsets of patients: patients with acute kidney injury (AUC = 0.875 for ANN, vs. SAPS, AUC = 0.642), patients with subarachnoid hemorrhage (AUC = 0.958 for BN, vs. SAPS, AUC = 0.84), and elderly patients undergoing open heart surgery (AUC = 0.94 for ANN, vs. EuroSCORE, AUC = 0.648). Rather than developing models with good external validity by including a heterogeneous patient population, an alternative approach would be to build models for specific patient subsets using one’s local database.