Predicting hospital mortality for patients in the intensive care unit: a comparison of artificial neural networks with logistic regression models.

OBJECTIVE: Logistic regression (LR), commonly used for hospital mortality prediction, has limitations. Artificial neural networks (ANNs) have been proposed as an alternative. We compared the performance of these approaches by using stepwise reductions in sample size. DESIGN: Prospective cohort study. SETTING: Seven intensive care units (ICU) at one tertiary care center. PATIENTS: Patients were 1,647 ICU admissions for whom first-day Acute Physiology and Chronic Health Evaluation III variables were collected. INTERVENTIONS: None. MEASUREMENTS AND MAIN RESULTS: We constructed LR and ANN models on a random set of 1,200 admissions (development set) and used the remaining 447 as the validation set. We repeated model construction on progressively smaller development sets (800, 400, and 200 admissions) and retested on the original validation set (n = 447). For each development set, we constructed models from two LR and two ANN architectures, organizing the independent variables differently. With the 1,200-admission development set, all models had good fit and discrimination on the validation set, where fit was assessed by the Hosmer-Lemeshow C statistic (range, 10.6-15.3; p > or = .05) and standardized mortality ratio (SMR) (range, 0.93 [95% confidence interval, 0.79-1.15] to 1.09 [95% confidence interval, 0.89-1.38]), and discrimination was assessed by the area under the receiver operating characteristic curve (range, 0.80-0.84). As development set sample size decreased, model performance on the validation set deteriorated rapidly, although the ANNs retained marginally better fit at 800 (best C statistic was 26.3 [p = .0009] and 13.1 [p = .11] for the LR and ANN models). Below 800, fit was poor with both approaches, with high C statistics (ranging from 22.8 [p <.004] to 633 [p <.0001]) and highly biased SMRs (seven of the eight models below 800 had SMRs of <0.85, with an upper confidence interval of <1). Discrimination ranged from 0.74 to 0.84 below 800. CONCLUSIONS: When sample size is adequate, LR and ANN models have similar performance. However, development sets of < or = 800 were generally inadequate. This is concerning, given typical sample sizes used for individual ICU mortality prediction.

An artificial neural network as a model for prediction of survival in trauma patients: validation for a regional trauma area.

BACKGROUND: To develop and validate an artificial neural network (ANN) for predicting survival of trauma patients based on standard prehospital variables, emergency room admission variables, and Injury Severity Score (ISS) using data derived from a regional area trauma system, and to compare this model with known trauma scoring systems. PATIENT POPULATION: The study was composed of 10,609 patients admitted to 24 hospitals comprising a seven-county suburban/rural trauma region adjacent to a major metropolitan area. The data was generated as part of the New York State trauma registry. Study period was from January 1993 through December 1996 (1993-1994: 5,168 patients; 1995: 2,768 patients; 1996: 2,673 patients). METHODS: A standard feed-forward back-propagation neural network was developed using Glasgow Coma Scale, systolic blood pressure, heart rate, respiratory rate, temperature, hematocrit, age, sex, intubation status, ICD-9-CM Injury E-code, and ISS as input variables. The network had a single layer of hidden nodes. Initial network development of the model was performed on the 1993-1994 data. Subsequent models were generated using the 1993, 1994, and 1995 data. The model was tested first on the 1995 and then on the 1996 data. The ANN model was tested against Trauma and Injury Severity Score (TRISS) and ISS using the receiver operator characteristic (ROC) area under the curve [ROC-A(z)], Lemeshow-Hosmer C-statistic, and calibration curves. RESULTS: The ANN showed good clustering of the data, with good separation of nonsurvivors and survivors. The ROCA(z) was 0.912 for the ANN, 0.895 for TRISS, and 0.766 for ISS. The ANN exceeded TRISS with respect to calibration (Lemeshow-Hosmer C-statistic: 7.4 for ANN; 17.1 for TRISS). The prediction of survivors was good for both models. The ANN exceeded TRISS in nonsurvivor prediction. CONCLUSION: An ANN developed for trauma patients using prehospital, emergency room admission data, and ISS gave good prediction of survival. It was accurate and had excellent calibration. This study expands our previous results developed at a single Level I trauma center and shows that an ANN model for predicting trauma deaths can be applied across hospitals with good results

Survival in patients with severe adult respiratory distress syndrome treated with high-level positive end-expiratory pressure.

OBJECTIVE: To assess the mortality rate and complications in a population of surgical patients with severe adult respiratory distress syndrome (ARDS) treated with positive end-expiratory pressure (PEEP) of > 15 cm H2O in an attempt to reduce intrapulmonary shunt to approximately 0.20 and reduce FIO2 to < 0.50. DESIGN: Retrospective review of patients treated by a standardized ventilatory support protocol at the time of their illness. SETTING: A 24-bed surgical intensive care unit in a university medical center. PATIENTS: All patients admitted to the surgical intensive care unit during a 34-month period who met the criteria for severe ARDS (Pao2 of < or = 70 torr [< or = 9.3 kPa] on an FIO2 of > or = 0.50, diffuse interstitial and/or alveolar infiltrates on chest radiograph, decreased lung compliance, no evidence of congestive heart failure, and a likely predisposing etiology) were evaluated. Patients treated with PEEP of > 15 cm H2O were selected for this review. INTERVENTIONS: Patients were treated by a protocol to achieve oxygenation end points, which consisted of maintaining arterial oxyhemoglobin saturation (as determined by pulse oximetry of > or = 0.92), while reducing FIO2 to < 0.50 and decreasing intrapulmonary shunt fraction to < or = 0.20 by adding PEEP. With the exception of patients with suspected intracranial hypertension related to closed-head injury, low-rate intermittent mandatory ventilation was the primary mode of ventilation. Pressure-support ventilation was added, when needed, to improve patient comfort, enhance spontaneous tidal volume, or improve CO2 excretion. MEASUREMENTS AND MAIN RESULTS: Eighty-six patients with severe ARDS were treated with a PEEP of > 15 cm H2O. Nineteen of these patients died early of severe closed-head injury or massive uncontrollable hemorrhage and were excluded from the evaluation. The remaining 67 patients had a mean Lung Injury Score of 3.3 during their treatment with high PEEP. Twenty (30%) of 67 patients died. Eight of the deaths occurred after decrease of ventilatory support and with acceptable blood gases. The other 12 patients who died had continued oxygenation deficits and received increased levels of ventilatory support at the time of death. Twenty-six (39%) of 67 patients had radiographic manifestations of barotrauma (pneumothorax, subcutaneous emphysema, etc.) related to their primary injuries or to complications related to central venous catheter placement. Seven (17%) of 41 patients developed clinical or radiographic signs of barotrauma while receiving high-level PEEP. The hemodynamic effects of increased airway pressure were managed with fluids and inotropic agents, when necessary, and did not limit the application of PEEP to reach the defined end point of treatment. CONCLUSIONS: This subset of patients with severe ARDS treated with high-level PEEP had a mortality rate lower than those rates previously reported by other researchers using more conventional ventilatory support and resuscitation techniques. FIO2 may be significantly reduced and PaO2 may be maintained at acceptable values by decreasing intrapulmonary shunt fraction using high-level PEEP.

High-level positive end-expiratory pressure management in trauma-associated adult respiratory distress syndrome.

This study evaluated the effect of high-level positive end-expiratory pressure (PEEP) on mortality, barotrauma, intrapulmonary shunt (Qsp/Qt), and oxygen delivery (DO2) in posttraumatic adult respiratory distress syndrome (ARDS). All hypoxemic trauma patients admitted to the surgical intensive care unit (SICU) in 1989-1990 who received PEEP greater than 15 cm H2O were included. The PEEP was titrated to achieve an intrapulmonary shunt (Qsp/Qt) of approximately 0.20, and FIO2 was weaned to less than 0.50. Hemodynamic and pulmonary variables at four distinct intervals were recorded. Fifty-nine patients received PEEP greater than 15 cm H2O. Of these, 19 patients died of severe head injury or uncontrollable hemorrhage (16 within 48 hours). Forty (29 male, 11 female) were evaluated in detail. The PEEP levels ranged from 18-50 cm H2O with a mean of 27. PaO2/FIO2 ratios and Qsp/Qt improved as PEEP therapy was titrated. Cardiac index and oxygen delivery were maintained or improved throughout PEEP therapy by transfusion and fluid resuscitation, with a mean maximum positive fluid balance of 21.1 L and an average of 51 units of blood and blood products transfused per patients during their SICU stay. Twenty-nine (73%) had evidence of barotrauma, the majority being pneumothoraces clearly related to the initial trauma. Only three (7.5%) had evidence of barotrauma not related to trauma or line insertion. Eight of 40 patients (20%) died. Mean ISS and RTS for the entire group were 32 and 5.88, respectively. We conclude that titration of PEEP to achieve a Qsp/Qt of approximately 0.20 is an attainable goal. This was accomplished with minimal hemodynamic effects or barotrauma and a low mortality rate.

Age-related superior mesenteric arterial flow changes in piglets: effects of feeding and hemorrhage.

Cardiovascular (CV) responses to feeding before and after 15% blood volume hemorrhage (H) were studied in lightly anesthetized piglets less than or equal to 2-days and greater than 2-wk-old. Superior mesenteric (Mes) and renal (Ren) arterial flows (F) were registered by electromagnetic probes before and continuously for 2 h after milk feeding (26 ml/kg) by gavage. Postprandially, in piglets less than or equal to 2-days-old MesF tended to increase (maximum change, mean +/- SE, 16.0 +/- 4.7%) at 30 min, whereas Mes vascular resistance (R) significantly decreased at 30, 90, and 120 min (17.3 +/- 6.2%). In piglets greater than 2-wk-old, MesF significantly increased by 30 min, which lasted 120 min (37.7 +/- 11.7%); MesR decreased by 20.0 +/- 5.8% at 90-120 min. Compared with the less than or equal to 2 days olds, the older piglets demonstrated greater and more sustained postprandial MesF increase. After H, regional F and pulse pressure (PP) decreased, heart rate and R increased in both groups. After stabilization, feeding induced insignificant CV changes in less than or equal to 2-day-old piglets. In contrast, increased MesF (44 +/- 14.4%) and PP (24 +/- 8.5%) and decreased MesR (31 +/- 8.9%) were observed in the greater than 2-wk-olds. MesF changes differed significantly between the age groups. The Ren vascular bed showed no consistent response in both age groups. Thus Mes vascular responses to feeding with or without preceding H in developing piglets were age related.

Age-related responses to stimulation of cardiopulmonary receptors in swine.

Cardiovascular responses to stimulation of cardiopulmonary receptors were evaluated in standard breed developing swine (1-2 days, 1-2 wk, and 2 mo of age) and sexually mature miniswine anesthetized with 0.25-0.50% halothane in O2. Cryptenamine, a mixture of veratrum viride alkaloids (VVA), was administered as right atrial or left ventricular bolus injections in doses of 5-20 micrograms/kg. In developing swine, mean aortic pressure (AoP) decreased after 10-20 micrograms/kg VVA in 1- to 2-wk olds and after 5-20 micrograms/kg VVA in 2-mo olds. Bradycardia was always elicited. Renal (Ren) and femoral (Fem) vasodilation occurred in the 1- to 2-wk-old and 2-mo-old groups after 5-20 micrograms/kg VVA. Mesenteric (Mes) vasodilation was elicited with 10-20 micrograms/kg VVA in the 1- to 2-wk-old group and with 5-20 micrograms/kg VVA in 2-mo olds. In the 1- to 2-day-old group, VVA did not significantly alter Ren, Fem, and Mes vascular resistance. In mature miniswine, 20 micrograms/kg VVA elicited decreases in AoP and heart rate that were similar in magnitude to responses obtained with only 5 micrograms/kg VVA in dogs. Cardiovascular responses to VVA were abolished after bilateral vagotomy but were not altered after denervation of the carotid sinuses. The results indicate postnatal maturation of the Bezold-Jarisch reflex in swine.