Each year approximately 3000 patients are admitted to the intensive care unit (ICU) in the University Medical Center Groningen (UMCG). In-hospital mortality of patients with emergency admission approaches 25%. Predicting outcome in the first hours after ICU admission, however, remains a challenge.

An vast amount of scoring systems has been developed for mortality prediction. Well known models, such as the LODS, MODS, CCI, SOFA, ODIN and the different generations of the APACHE, MPM and SAPS, are increasingly compared with new models, such as the SICULA, ICNARC, ANZROD and SMS-ICU. The predictive value of scoring systems deteriorates over time due to changes in patient characteristics and treatment, making it crucial to update existing models or develop new models. Other reasons given for the need of models are the complexity and lack of availability of variables in some of the existing scoring systems, the better discriminating value while using simple, standardly measured variables, and the limited generalizability of some scoring systems in different patient populations. Not only are simple systems (such as the CIS and SMS-ICU) found to be at least as predictive for mortality as complex models such as the APACHE IV, but, while using simplified systems, mortality can also be reasonably predicted within only a few hours after admission. Both simplicity and the potential to predict mortality shortly after admission increase the usability, and consequently the reliability, of those prediction models. This increases the potential of those models to be used in practice.

Most studies however compare only two to four models in their patient population and lack in their description of the performance of the different models. Parameters necessary to compare the performance of models are at least calibration, discrimination, negative predicting value, positive predicting value, sensitivity and specificity. Lacking an adequate description of the performance of the model limits to what extent the study can be used to compare models in different populations. Thus, all usable models should be compared with newly build models, and the performance of the different models should be extensively described to allow comparison of the models.

Not only models based on simple, readily available variables available within hours after admission are promising, but also the concept of combining measurements straight after ICU admission with information on the course of illness. It is likely that the course of a variable over time is more indicative than a static measurement. This study will provide a structure in which every patient admitted to the ICU will be investigated and included within 3 hours and after 12 hours after admission, making longitudinal measurements and various add-on studies possible. Longitudinal measurements are the first example of an add-on study; another example is the capability of nurses and physicians to predict outcome. Current evidence suggests that physicians might predict mortality more accurately than scorings systems. This finding may, however, be highly biased, since at least physicians play a major role in end-of-life decision making. More recent studies also focus on the accuracy of nurses in predicting mortality, with diverse outcomes. The role of other health care professionals, like residents and students, remain to be studied.

Implementing a systematic data collection process is the first step towards making data-driven research possible, a growing need in medical disciplines such as critical care, which requires increasingly more accurate prognostic models. Therefore, the aim of this study is to systematically collect data of all selected variables, thus minimizing incompleteness, and allowing for the calculation of mortality prediction scores according to currently available mortality or severity of disease prediction models. Moreover, during investigation reliability of measurements could be checked for validity. This creates the possibility to compare the performance of all models in one population and identify models which are useful to predict severity of disease. A registry will be created with this primary objective which also provides the opportunity to start multiple ''add-on'' studies for specific research questions. Examples of add-on studies are 1) the association between time-dependent variables which are longitudinally measured, and mortality/acute and chronic co-morbidity, 2) the association between fluid status and acute kidney injury, and 3) not only the capability of the treating physician to predict mortality, but also the capability of the nurses, residents and students to do so.

Purpose:

The purpose of this study is to expand the infrastructure for a registry with longitudinal and repeated measurements, shortly after admittance, which is flexible to incorporate temporarily added specific research questions on the outcome of critically ill patients.

Registry procedures:

Eligible patients will be included within 3 hours after their arrival on the Intensive Care Unit. After inclusion all study parameters will be obtained through physical and physiological examination. This will be repeated at 12 circa hours, and circa 24 hours thereafter.

Monitoring:

Monitoring will be performed by independent researchers of the department of critical care of the University Medical Center Groningen (UMCG). Audits are planned to take place once a year.

Recruitment:

Inclusion of patients and measurements of variables will be performed by the study coordinator or a co-researcher under supervision and responsibility of the principle investigator. Due to its observational nature, informed consent will not be obtained for this study.

Source data verification:

At inclusion all conventional hemodynamic variables are derived by physical examination and recording data from the basic hemodynamic monitoring (Philips ImageVue monitor with tracing of heart rate, electrocardiogram (ECG), SpO2, arterial pressure from arterial line pressure measurement and/or from non-invasively blood pressure monitoring). All variables are predefined (see data dictionary) to standardize all measurements by student researchers.

General patient characteristics and laboratory measurements were recorded from electronic patient charts and the Acute Physiology and Chronic Health Evaluation (APACHE) II and IV, Simplified Acute Physiology Score II (SAPS) scores are extracted from our local National Intensive Care Evaluation (NICE) database. Follow-up of all-cause mortality at 90 days and 3 years is acquired using the municipal personal records database. Long-term follow-up is gathered through various resources, such as hospital records, family physicians and/or direct contact.

Data collection:

Within 3 hours of ICU admission, all variables will be obtained through a onetime clinical examination variables and nurses and physicians will be asked for their prognostications. Other variables (i.e. laboratory and imaging values) will be obtained from the electronic patient charts. The rationale and specific details of measuring each variable are described extensively below.

Systemic circulatory variables:

Heart rate (HR): it will be recorded from the bedside electrocardiographic monitor. In case of an irregular rhythm (i.e. atrial fibrillation) the investigators will use the mean heart rate over a minute. Apart from heart rate the presence of atrial fibrillation will be recorded.

Systolic blood pressure (SBP), diastolic blood pressure (DBP) and mean arterial pressure (MAP): these will be obtained by intravascular measurement using an arterial line, which is part of usual care. If no arterial line inserted, the non-invasive measurement will be collected.

Central venous pressure (CVP): this will be recorded in case a central venous line is present in the internal jugular or subclavian vein.

Micro- and peripheral circulatory variables:

Capillary refill time (CRT): this will be measured after 10 seconds of exerting firm pressure preferably on the distal phalanx of the index finger and on the central part of the knee. No cutoff score will be used, but a continuous measurement because of the discrepancy in values considered normal by different researches for CRT.

Skin temperature (Tskin): this will be measured subjectively and objectively. The subjective measure will be conducted by palpating the patient's extremities. A distinction between either 'warm' or 'cold' will be made using the dorsal surface of the hands of the examiner. Patients will be considered to have 'cold' skin extremities if all examined extremities are considered cool, or if only the lower extremities are cool despite warm upper extremities.

To objectify the skin temperature, the use of a central-to-peripheral and peripheral-to-ambient temperature difference (respectively dTc-p and dTp-a) or the forearm-to-finger skin-temperature gradient (Tskin-diff) have been proposed in literature. The investigators will make use of the central-to-peripheral measurements:

Central-to-peripheral temperature difference (dTc-p): to measure this difference the investigators will compare bladder temperature, as measured by a bladder thermistor catheter, with foot temperature, measured by a skin probe (DeRoyal Skin Temperature Sensor product nr 81-010400EU) on either the left or the right foot (dorsum). The investigators will use bladder temperature as a surrogate for central temperature, and toe temperature as a peripheral measure. In literature a temperature difference of either 5°C or 7 °C is generally used as an upper limit. The investigators will therefore regard values higher than 7°C as abnormal.

The mottling score: this score was described by Ait-Oufella et al in 2011. Mottling is the patchy discoloration of the skin caused by microcirculatory dysfunction. It usually involves the area around the knee. The Mottling score ranges from 0 to 5, depending on the extensiveness of the mottled area. A score of 0 or 1 is regarded mild, 2 or 3 moderate and 4 or 5 severe.

Urine output (ml/kg/h): this is also measured as part of regular care. The investigators will use the total urine output before examination (i.e. from admission to examination). In patients with pre-existing renal failure the urinary output will not be used.

Other variables:

Respiratory rate: this will be recorded of the bedside electrocardiographic monitor. If a patient is on mechanical ventilation, see below.

Mechanical ventilation: data on the presence and type of mechanical ventilation will be gathered, as well as basic information on respiratory conditions (e.g. PEEP and FiO2).

Inotropic and vasopressor use: any inotrope or vasopressin requirement, type, dose and speed will be recorded. Considered as inotropic/vasopressors are noradrenalin, vasopressin, dobutamine, dopamine and milrinone. Considered as sedative are midazolam, propofol and S-ketamine.

Estimations of pump function and peripheral circulation: an estimation will be made, either by a member of the treating team, or by the researcher.

Serum lactate, hemoglobin, troponin-T and creatinine: these are determined as part of regular care. For study purposes the investigators will use the value closest to our examination.

Other biochemical values determined as part of regular care, such as pH, leucocytes, hematocrit, et cetera will also be recorded.

After the clinical examination has been performed, information on the following general characteristics will be extracted from the NICE database. This includes demographic data (such as BMI, age and sex), medical history, diagnoses and severity of illness as evaluated by the different generations of the APACHE scores, SAPS, the Sequential Organ Failure Assessment (SOFA) and all other prediction models published. To this end, an extensive literature search will be created and published and thereafter performed to ensure inclusion of all published models with their variables. Furthermore, the investigators will collect laboratory values (details are described above), urine output (details are described above) and routine admission ECG's and imaging. After 30 days the investigators will assess the patient files again to gather information on total ICU stay in days.

Data management:

Data will be recorded using OpenClinica and transferred for analysis. After transfer from OpenClinica, all data will be managed in a database created using Stata version 15 (StataCorp, College Station, TX). All study subjects will receive a study subject ID, compiled of the study name and their inclusion number. This study subject ID will be used in both OpenClinica and Stata. Only a researcher with 'study director' account properties in OpenClinica will be able to link study subject ID to patient number.

Sample size assessment:

Each year 3000 patients are admitted to one of four ICU units. Approximately 1500 of these admissions are unplanned emergency admissions. The investigators estimate that half of these unplanned admissions fulfill the inclusion criteria. This leaves 750 patients eligible for inclusion. However, the investigators assume that they will not be able to include all eligible patients for logistic and practical reasons. Therefore, the investigators aim to include 400 patients per year.

Plan for missing data:

First, the pattern underlying our missing data will be explored: missing completely at random, missing at random, or missing not at random. Based on previous experience with the SICS-I, it is expected that data will be missing at random. Therefore, primary analyses will be conducted with imputation for missing data using multiple imputations. Robustness of conclusions will be checked by secondary sensitivity analyses only including available data.

Statistical analysis plan:

The investigators will use the general characteristics to create a baseline table. Statistical analyses will be performed using Stata version 15 (StataCorp, College Station, TX). Data will be presented as means with standard deviation if normally distributed or as medians with ranges in case of skewed data.

Univariate analyses will be conducted and all variables with p<0.1 will be included in the multivariate models. Multivariate analyses will be conducted using a stepwise model. Cardiac output will be modeled using linear regression and mortality will be modeled using logistic regression. All analyses will be adjusted for age and gender; other general characteristics will not be added to the model standardly. All analyses will be tested two-sided and p-values of less than 0.05 will be considered statistical significant.

If sample size permits, the investigators will conduct an analysis in different subpopulations. Examples of subpopulations that may be eligible for further analysis are those with sepsis and or severe circulatory shock.

Future plans include the use of Machine Learning methods for further data analysis. The SOCCS has been designed so that data collection is simultaneously as uniform and as vast as possible. As previously described, almost all patients admitted to the ICU will be included, thanks to the availability of a trained student team on-call 24 hours a day. This guarantees that new patients can be included at any time, and provides the possibility of creating a large set of uniform data. Many variables are registered accurately at bedside for all patients acutely admitted to the ICU, within 3 hours of admission. Both more traditional as well as innovative data analysis projects benefit from an optimal data collection process. All data collected for the SOCCS study will be compiled in an extensive and complete database, from which Machine Learning-based predictive modeling studies can easily be defined and set up. At a later stage, these data will be used for building predictive or risk-stratification models, which aim at probabilistically tracking the evolution of critically ill patients or at highlighting potentially interesting parameters for subsequent studies.

Machine Learning (ML) is a branch of Artificial Intelligence which allows data scientists to design supervised or unsupervised algorithms to "learn" from generally large data samples by means of inference. ML started as a well-known method in gene analysis, having since then spread to multiple other fields of medicine. Its potential to boost clinically-oriented research has been shown, primarily in identifying and prioritizing specific components of care, the study of which may bring the most benefit to patients. Furthermore, the use of ML-based frameworks in critical care has been reported by several authors, using different types of data. These data can be obtained in different ways, with the most common being through bedside measurements, such as performed for the SOCCS and SICS studies, or by collecting a larger number of parameters from prior physiologic data from a group of similar subjects.

Inclusion Criteria:

• Emergency admission
• Expected stay > 24 hours

Exclusion Criteria:

• Age < 18 years
• Planned admission either after surgery or for other reasons
• Suicide attempts due to acute psychiatric 'derailment', mental retardation or a language barrier

The study will be conducted in the department of Critical Care of the University of Groningen, University Medical Center Groningen, a tertiary teaching hospital in the Northern part of the Netherlands.

According guidelines for sharing.