Using data mining for prediction of hospital length of stay: an application of the CRISP-DM Methodology

Hospitals are nowadays collecting vast amounts of data related with patient records. All this data hold valuable knowledge that can be used to improve hospital decision making. Data mining techniques aim precisely at the extraction of useful knowledge from raw data. This work describes an implementation of a medical data mining project approach based on the CRISP-DM methodology. Recent real-world data, from 2000 to 2013, were collected from a Portuguese hospital and related with inpatient hospitalization. The goal was to predict generic hospital Length Of Stay based on indicators that are commonly available at the hospitalization process (e.g., gender, age, episode type, medical specialty). At the data preparation stage, the data were cleaned and variables were selected and transformed, leading to 14 inputs. Next, at the modeling stage, a regression approach was adopted, where six learning methods were compared: Average Prediction, Multiple Regression, Decision Tree, Artificial Neural Network ensemble, Support Vector Machine and Random Forest. The best learning model was obtained by the Random Forest method, which presents a high quality coefficient of determination value (0.81). This model was then opened by using a sensitivity analysis procedure that revealed three influential input attributes: the hospital episode type, the physical service where the patient is hospitalized and the associated medical specialty. Such extracted knowledge confirmed that the obtained predictive model is credible and with potential value for supporting decisions of hospital managers.

White-coat hypertension and normotension in the League of Hypertension of the Hospital das Clínicas, FMUSP: prevalence, clinical and demographic characteristics

OBJECTIVE: To assess the prevalence of white-coat normortension, white-coat hypertension, and white-coat effect. METHODS: We assessed 670 medical records of patients from the League of Hypertension of the Hospital das Clínicas of the Medical School of the University of São Paulo. White-coat hypertension (blood pressure at the medical office: mean of 3 measurements with the oscillometric device ³140 or ³90 mmHg, or both, and ambulatory blood pressure monitoring mean during wakefulness < 135/85) and white-coat normotension (office blood pressure < 140/90 and blood pressure during wakefulness on ambulatory blood pressure monitoring ³ 135/85) were analyzed in 183 patients taking no medication. The white-coat effect (difference between office and ambulatory blood pressure > 20 mmHg for systolic and 10 mmHg for diastolic) was analyzed in 487 patients on treatment, 374 of whom underwent multivariate analysis to identify the variables that better explain the white-coat effect. RESULTS: Prevalence of white-coat normotension was 12%, prevalence of white-coat hypertension was 20%, and prevalence of the white-coat effect was 27%. A significant correlation (p<0.05) was observed between white-coat hypertension and familial history of hypertension, and between the white-coat effect and sex, severity of the office diastolic blood pressure, and thickness of left ventricular posterior wall. CONCLUSION: White-coat hypertension, white-coat normotension, and white-coat effect should be considered in the diagnosis of hypertension.

Acute myocardial infarction: clinical and epidemiological profile and factors associated with in-hospital death in the municipality of Rio de Janeiro

OBJECTIVE: To study the factors associated with the risk of in-hospital death in acute myocardial infarction in the Brazilian public health system in Rio de Janeiro, Brazil. METHODS: Sectional study of a sample with 391 randomly drawn medical records of the hospitalizations due to acute myocardial infarction recorded in the hospital information system in 1997. RESULTS: The diagnosis was confirmed in 91.7% of the cases; 61.5% males; age = 60.2 ± 2.4 years; delta time until hospitalization of 11 hours; 25.3% were diabetic; 58.1% were hypertensive; 82.6% were in Killip I class. In-hospital mortality was 20.6%. Thrombolysis was used in 19.5%; acetylsalicylic acid (ASA) 86.5%; beta-blockers 49%; angiotensin-converting enzyme (ACE) inhibitors 63.3%; calcium channel blockers 30.5%. Factors associated with increased death: age (61-80 years: OR=2.5; > 80 years: OR=9.6); Killip class (II: OR=1.9; III: OR=6; IV: OR=26.5); diabetes (OR=2.4); ventricular tachycardia (OR=8.5); ventricular fibrillation (OR=34); recurrent ischemia (OR=2.7). The use of ASA (OR=0.3), beta-blockers (OR=0.3), and ACE inhibitors (OR=0.4) was associated with a reduction in the chance of death. CONCLUSION: General lethality was high and some interventions of confirmed efficacy were underutilizated. The logistic model showed the beneficial effect of beta-blockers, and ACE inhibitors on the risk of in-hospital death.

Level of Physical Activity and In-Hospital Course of Patients with Acute Coronary Syndrome

Abstract Background: Acute coronary syndrome (ACS) is one of the main causes of morbidity and mortality in the modern world. A sedentary lifestyle, present in 85% of the Brazilian population, is considered a risk factor for the development of coronary artery disease. However, the correlation of a sedentary lifestyle with cardiovascular events (CVE) during hospitalization for ACS is not well established. Objective: To evaluate the association between physical activity level, assessed with the International Physical Activity Questionnaire (IPAQ), with in-hospital prognosis in patients with ACS. Methods: Observational, cross-sectional, and analytical study with 215 subjects with a diagnosis of ACS consecutively admitted to a referral hospital for cardiac patients between July 2009 and February 2011. All volunteers answered the short version of the IPAQ and were observed for the occurrence of CVE during hospitalization with a standardized assessment conducted by the researcher and corroborated by data from medical records. Results: The patients were admitted with diagnoses of unstable angina (34.4%), acute myocardial infarction (AMI) without ST elevation (41.4%), and AMI with ST elevation (24.2%). According to the level of physical activity, the patients were classified as non-active (56.3%) and active (43.7%). A CVE occurred in 35.3% of the cohort. The occurrence of in-hospital complications was associated with the length of hospital stay (odds ratio [OR] = 1.15) and physical inactivity (OR = 2.54), and was independent of age, systolic blood pressure, and prior congestive heart failure. Conclusion: A physically active lifestyle reduces the risk of CVE during hospitalization in patients with ACS.

Informatització de l'arxiu d'històries clíniques d'un hospital

El sector sanitari és un dels més importants i sensibles dintre de l’economia espanyola, ja que és un element clau del benestar social. Per aquest motiu, el sector de la sanitat s’enfronta a nombrosos reptes, i les Tecnologies de la Informació i les Comunicacions (TIC) obren un ampli ventall de possibilitats de renovació i millora. És per això que Hewlett-Packard està desenvolupant una aplicació informàtica per a hospitals, anomenada HCIS (Health Care Information System). El projecte que es presenta en aquesta memòria és el desenvolupament del mòdul d’HCIS per a gestionar l’Arxiu d’Històries Clíniques d’un hospital, i la seva implantació a la Corporació Sanitària Parc Taulí de Sabadell.

Is hospital discharge administrative data an appropriate source of information for cancer registries purposes? Some insights from four Spanish registries

Background. The use of hospital discharge administrative data (HDAD) has been recommended for automating, improving, even substituting, population-based cancer registries. The frequency of false positive and false negative cases recommends local validation. Methods. The aim of this study was to detect newly diagnosed, false positive and false negative cases of cancer from hospital discharge claims, using four Spanish population-based cancer registries as the gold standard. Prostate cancer was used as a case study. Results. A total of 2286 incident cases of prostate cancer registered in 2000 were used for validation. In the most sensitive algorithm (that using five diagnostic codes), estimates for Sensitivity ranged from 14.5% (CI95% 10.3-19.6) to 45.7% (CI95% 41.4-50.1). In the most predictive algorithm (that using five diagnostic and five surgical codes) Positive Predictive Value estimates ranged from 55.9% (CI95% 42.4-68.8) to 74.3% (CI95% 67.0-80.6). The most frequent reason for false positive cases was the number of prevalent cases inadequately considered as newly diagnosed cancers, ranging from 61.1% to 82.3% of false positive cases. The most frequent reason for false negative cases was related to the number of cases not attended in hospital settings. In this case, figures ranged from 34.4% to 69.7% of false negative cases, in the most predictive algorithm. Conclusions. HDAD might be a helpful tool for cancer registries to reach their goals. The findings suggest that, for automating cancer registries, algorithms combining diagnoses and procedures are the best option. However, for cancer surveillance purposes, in those cancers like prostate cancer in which care is not only hospital-based, combining inpatient and outpatient information will be required.

Hospital volume and patient outcomes in pulmonary embolism.

BACKGROUND: In numerous high-risk medical and surgical conditions, a greater volume of patients undergoing treatment in a given setting or facility is associated with better survival. For patients with pulmonary embolism, the relation between the number of patients treated in a hospital (volume) and patient outcome is unknown. METHODS: We studied discharge records from 186 acute care hospitals in Pennsylvania for a total of 15 531 patients for whom the primary diagnosis was pulmonary embolism. The study outcomes were all-cause mortality in hospital and within 30 days after presentation for pulmonary embolism and the length of hospital stay. We used logistic models to study the association between hospital volume and 30-day mortality and discrete survival models to study the association between in-hospital mortality and time to hospital discharge. RESULTS: The median annual hospital volume for pulmonary embolism was 20 patients (interquartile range 10-42). Overall in-hospital mortality was 6.0%, whereas 30-day mortality was 9.3%. In multivariable analysis, very-high-volume hospitals (> or = 42 cases per year) had a significantly lower odds of in-hospital death (odds ratio [OR] 0.71, 95% confidence interval [CI] 0.51-0.99) and of 30-day death (OR 0.71, 95% CI 0.54-0.92) than very-low-volume hospitals (< 10 cases per year). Although patients in the very-high-volume hospitals had a slightly longer length of stay than those in the very-low-volume hospitals (mean difference 0.7 days), there was no association between volume and length of stay. INTERPRETATION: In hospitals with a high volume of cases, pulmonary embolism was associated with lower short-term mortality. Further research is required to determine the causes of the relation between volume and outcome for patients with pulmonary embolism.

Prevalence of diabetes in a cancer population in a Malaga hospital.

BACKGROUND There are multiple risk factors for cancer, including obesity, sedentary lifestyle, diabetes (DM). Hormon Insulin is a growth factor that promotes cellular differentiation. AIMS The aim of our study is to observe impaired glycaemia in cancer population compared with control. METHODS We studied the prevalence of diabetes (DM) and impaired fasting glycaemia (IFG) in 374 patients with different types of cancer before treatment, by medical records in a Malaga hospital (Spain). We compared the prevalence of basal hyperglycaemia in these patients with general population, within an age range and by gender. RESULTS AND DISCUSSION The prevalence of diabetes was 32.35% in our cancer patients. The comparison depends of age range, and by gender prevalence was: 45-54 years, DM: 40.91% in men cases, versus (vs.) 14.5% in men control (p = 0.005). 55-64 years, IFG: 23.08% in women cases, vs. 5.9% in women control (p = 0.001). 65-74 years, DM: 47.13% in men cases, vs. 25.4% in men control (p = 0.000), and IFG: 23.81% in women cases, vs. 9.5% in women control (p = 0.019). We found a higher prevalence of diabetes in specific types of cancer such as prostate (p < 0.005). Moreover, men had a higher prevalence of diabetes or less diabetes control than women in our cancer sample. CONCLUSIONS We recommend an OGTT (oral glucose tolerance test) for better diagnosis of possible DM in patients with cancer, and an appropriate treatment. It may be an independent risk factor for cancer to have decreased insulin activity, or DM.

Validation of the potentially avoidable hospital readmission rate as a routine indicator of the quality of hospital care

BACKGROUND: The hospital readmission rate has been proposed as an important outcome indicator computable from routine statistics. However, most commonly used measures raise conceptual issues. OBJECTIVES: We sought to evaluate the usefulness of the computerized algorithm for identifying avoidable readmissions on the basis of minimum bias, criterion validity, and measurement precision. RESEARCH DESIGN AND SUBJECTS: A total of 131,809 hospitalizations of patients discharged alive from 49 hospitals were used to compare the predictive performance of risk adjustment methods. A subset of a random sample of 570 medical records of discharge/readmission pairs in 12 hospitals were reviewed to estimate the predictive value of the screening of potentially avoidable readmissions. MEASURES: Potentially avoidable readmissions, defined as readmissions related to a condition of the previous hospitalization and not expected as part of a program of care and occurring within 30 days after the previous discharge, were identified by a computerized algorithm. Unavoidable readmissions were considered as censored events. RESULTS: A total of 5.2% of hospitalizations were followed by a potentially avoidable readmission, 17% of them in a different hospital. The predictive value of the screen was 78%; 27% of screened readmissions were judged clearly avoidable. The correlation between the hospital rate of clearly avoidable readmission and all readmissions rate, potentially avoidable readmissions rate or the ratio of observed to expected readmissions were respectively 0.42, 0.56 and 0.66. Adjustment models using clinical information performed better. CONCLUSION: Adjusted rates of potentially avoidable readmissions are scientifically sound enough to warrant their inclusion in hospital quality surveillance.

Lessons learned after 2 full scale disaster exercises in a Swiss pediatric hospital

Introduction: Following a disaster, up to 50% of mass casualties are children. The number of disaster increases worldwide, including in Switzerland. Following national order, the mapping of the various risks of disaster in Switzerland will be completed by the end of 2012. Pre-hospital disaster drills and plans are well established and regularly tested. In-hospital disaster plans are much less frequently tested, if only available. Pediatric in-hospital full scale disaster exercises have never been reported in Switzerland. Based on our local constraints, we set up and evaluated a disaster plan during two full scale exercises. Methods: In a university hospital treating more than 35 000 pediatric emergencies per year, two exercises involving mock victims of a disaster aged 9 to 14 years old were successively set up over a period of 3 years. The exercises were planned during the day, without modification of the normal emergency room activities. The hospital staff was informed and trained in advance. Variables such as the alarm timing and transmission, triage set-up and function, special disaster medical records utilization, communication and victims' identification were assessed. Family members participated in the second exercise. An evaluation team observed and record exercises activities, identifying strength and weaknesses. Results: On two separate occasions, a total of 44 mock patients participated, were triaged, admitted and treated in the hospital according to usual standards of care. Alarm transmission was not appropriate during the first exercise. Triage overload occurred on one occasion. In-hospital communication needed readjustment. Identification and in-hospital tracking of the children remained problematic. Hospital employees showed great enthusiasm and stressed the positive effect of full scale exercises on their knowledge of the hospital disaster plan. Conclusions: Performing real life disaster exercises in a pediatric hospital was very beneficial. The disaster plan was adapted to local needs and updated accordingly. An alarm transmission protocol was elaborated and tested. Triage set-up was adapted and tested. A hospital identification plan for injured children was created and tested. Full scale hospital exercises evaluating disaster plans revealed several weaknesses in the system. Practice readjustments based on local experience were made. A tested pediatric disaster plan adapted to local constraints could minimize chaos, optimize care and support in the event of a real disaster. Children's identification and family reunification following a disaster remains a challenge.

Inpatient diabetes care in a Swiss multi-site hospital: is there room for improvement?

Background and objective: Optimal care of diabetic patients (DPs) decreases the risk of complications. Close blood glucose monitoring can improve patient outcomes and shorten hospital stay. The objective of this pilot study was to evaluate the treatment of hospitalized DPs according to the current standards, including their diabetic treatment and drugs to prevent diabetes related complications [=guardian drugs: angiotensin converting enzyme inhibitors (ACEI) or Angiotensin II Receptor Blockers (ARB), antiplatelet drugs, statins]. Guidelines of the American Diabetes Association (ADA) [1] were used as reference as they were the most recent and exhaustive for hospital care. Design: Observational pilot study: analysis of the medical records of all DPs seen by the clinical pharmacists during medical rounds in different hospital units. An assessment was made by assigning points for fulfilling the different criteria according to ADA and then by dividing the total by the maximum achievable points (scale 0-1; 1 = all criteria fulfilled). Setting: Different Internal Medicine and Geriatric Units of the (multi-site) Ho^pital du Valais. Main outcome measures: - Completeness of diabetes-related information: type of diabetes, medical history, weight, albuminuria status, renal function, blood pressure, (recent) lipid profile. - Management of blood glucose: Hb1Ac, glycemic control, plan for treating hyper-/hypoglycaemia. - Presence of guardian drugs if indicated. Results: Medical records of 42 patients in 10 different units were analysed (18 women, 24 men, mean age 75.4 ± 11 years). 41 had type 2 diabetes. - Completeness of diabetes-related information: 0.8 ± 0.1. Information often missing: insulin-dependence (43%) and lipid profile (86%). - Management of blood glucose: 0.5 ± 0.2. 15 patients had suboptimal glycemic balance (target glycaemia 7.2-11.2 mmol/ l, with values[11.2 or\3.8 mmol/l, or Hb1Ac[7%), 10 patients had a deregulated balance (more than 10 values[11.2 mmol/l or \3.8 mmol/l and even values[15 mmol/l). - Presence of guardian drugs if indicated: ACEI/ARB: 19 of 23 patients (82.6%), statin: 16 of 40 patients (40%), antiplatelet drug: 16 of 39 patients (41%). Conclusions: Blood glucose control was insufficient in many DPs and prescription of statins and antiplatelet drugs was often missing. If confirmed by a larger study, these two points need to be optimised. As it is not always possible and appropriate to make those changes during hospital stay, a further project should assess and optimise diabetes care across both inpatient and outpatient settings.

Proportion of out-of-hospital adult non-traumatic cardiac or respiratory arrest among calls for seizure.

OBJECTIVES: To measure the proportion of adult non-traumatic cardiac or respiratory arrest among calls for seizure to an emergency medical dispatch centre and to record whether known epileptic patients present cardiac or respiratory arrest together with seizure. METHODS: This 2-year prospective observational investigation involved the collection of tape recordings of all incoming calls to the emergency medical dispatch centre, in which an out-of-hospital non-traumatic seizure was the chief complaint in patients >18 years, in addition to the paramedics' records of all patients who presented with respiratory or cardiac arrest. The authors also recorded whether the bystander spontaneously mentioned to the dispatcher that the victim was known to have epilepsy. RESULTS: During the 24-month period, the call centre received 561 incoming calls for an out-of-hospital non-traumatic seizure in an adult. Twelve cases were classified as cardiac or respiratory arrest by paramedics. In one case, the caller spontaneously mentioned that the victim had a history of epilepsy. The proportion of cardiac or respiratory arrest among calls for seizure was 2.1%. CONCLUSION: Although these cases are rare, dispatchers should closely monitor seizure patients with the help of bystanders to exclude an out-of-hospital cardiac or respiratory arrest, in which case the dispatcher can offer telephone cardiopulmonary resuscitation advice until the paramedics arrive. Whenever the activity of the centre allows it and no new incoming call is on hold, this can be achieved by staying on the line with the caller or by calling back. A history of epilepsy should not modify the type of monitoring performed by the dispatcher as those patients may also have an arrest together with seizure.

Measuring potentially avoidable hospital readmissions.

The objectives of this study were to develop a computerized method to screen for potentially avoidable hospital readmissions using routinely collected data and a prediction model to adjust rates for case mix. We studied hospital information system data of a random sample of 3,474 inpatients discharged alive in 1997 from a university hospital and medical records of those (1,115) readmitted within 1 year. The gold standard was set on the basis of the hospital data and medical records: all readmissions were classified as foreseen readmissions, unforeseen readmissions for a new affection, or unforeseen readmissions for a previously known affection. The latter category was submitted to a systematic medical record review to identify the main cause of readmission. Potentially avoidable readmissions were defined as a subgroup of unforeseen readmissions for a previously known affection occurring within an appropriate interval, set to maximize the chance of detecting avoidable readmissions. The computerized screening algorithm was strictly based on routine statistics: diagnosis and procedures coding and admission mode. The prediction was based on a Poisson regression model. There were 454 (13.1%) unforeseen readmissions for a previously known affection within 1 year. Fifty-nine readmissions (1.7%) were judged avoidable, most of them occurring within 1 month, which was the interval used to define potentially avoidable readmissions (n = 174, 5.0%). The intra-sample sensitivity and specificity of the screening algorithm both reached approximately 96%. Higher risk for potentially avoidable readmission was associated with previous hospitalizations, high comorbidity index, and long length of stay; lower risk was associated with surgery and delivery. The model offers satisfactory predictive performance and a good medical plausibility. The proposed measure could be used as an indicator of inpatient care outcome. However, the instrument should be validated using other sets of data from various hospitals.

Epidemiology and injury patterns of patients consulting following assault by nightclub security agents in a university hospital emergency service and an associated specialised forensic consultation

Introduction: The Violence Medical Unit (VMU), a specialised forensic medical consultation, was created at the Lausanne university Hospital in 2006. All patients consulting at the ED for interpersonal violencerelated injury are referred to the VMU, which provides forensic documentation of the injury and referral to the relevant community based victim-support organisations within 48 hours of the ED visit. This frees the ED medical staff from forensic injury documentation and legal/social referral, tasks for which they lack both time and training. Among community violence, assaults by nightclub security agents against patrons have increased from 6% to 10% between 2007 and 2009. We set out to characterise the demographics, assault mechanisms, subsequent injuries, prior alcohol intake and ED & VMU costs incurred by this group of patients. Methods: We retrospectively included all patients consulting at the VMU due to assault by nightclub security agents from January 2007 to December 2009. Data was obtained from ED & VMU medical, nursing and administrative records. Results: Our sample included 70 patients, of which 64 were referred by the CHUV ED. The victims were typically young (median age 29) males (93%). 77% of assaults occurred on the weekend between 12 PM and 4 AM, and 73% of the victims were under the influence of alcohol. 83% of the patients were punched, kicked and/or head-butted; 9% had been struck with a blunt instrument. 80% of the injuries were in the head and neck area and 19% of the victims sustained fractures. 21% of the victims were prescribed medical leave. Total ED & VMU costs averaged 1048 SFr. Conclusion: Medical staff treating this population of assault victims must be aware of the assault mechanisms and injury patterns, in particular the high probability of fractures, in order to provide adequate diagnosis and care. Associated inebriation mandates liberal use of radiology, as delayed or missed diagnosis may have medical, medicolegal and legal implications. Emergency medical services play an important role in detecting and reporting of such incidents. Centralised management of the forensic documentation facilitates referral to victim support organisations and epidemiological data collection. Magnitudes and trends of the different types of violence can be determined, and this information can be then impact public safety management policies.

Incidence of uterine post abortion infection at Hospital de Clínicas de Porto Alegre. Is prophylactic antibiotic necessary?

OBJECTIVE: To identify the incidence of pelvic infection after miscarriage undergoing uterine evacuation in a tertiary hospital in southern Brazil and to compare with the international literature.METHODS: we reviewed electronic medical records of the Hospital de Clinicas de Porto Alegre of all patients who underwent uterine evacuation for miscarriage between August 2008 and January 2012 were reviewed. We included all patients submitted to uterine curettage due to abortion and who had outpatient visits for review after the procedure. We calculated emographic and laboratory data of the study population, number needed for treatment (NNT) and number needed to harm (NNH).RESULTS: of the 857 revised electronic medical records, 377 patients were subjected to uterine evacuation for miscarriage; 55 cases were lost to follow-up, leaving 322 cases that were classified as not infected abortion on admission. The majority of the population was white (79%); HIV prevalence and positive VDRL was 0.3% and 2%, respectively. By following these 322 cases for a minimum of seven days, it was found that the incidence of post-procedure infection was 1.8% (95% CI 0.8 to 4). The NNT and NNH calculated for 42 months were 63 and 39, respectively.CONCLUSION: The incidence of post-abortion infection between August 2008 to January 2012 was 1.8% (0.8 to 4).