Mortality prediction in the intensive care unit

Intensive care unit (https://www.roanemedical.com/wp-content/uploads/sites/10/2016/03/Intensive-Care-Unit.jpg)

Scoring systems to predict mortality in the intensive care unit have been used since the 1980s. A well-performing predictive model provides the hospital with valuable information about the patients condition and aids in patient management. The rising costs of critical care medicine (CCM) have created a high demand for an accurate system predicting patient outcome.

A precise measure of patient management is the bed occupancy rate (number of inpatient days / number of bed days available) in the ICU. Ideally, this percentage should be close to 100% or on the rise year-to-year.

Another comparison is the critical care costs vs. real GDP (adjusted for price changes). The percent increase should preferably be similar for both markers.

Data for the years 1985, 2000, 2005 and 2010 were available in publications by Neil A. Halpern, MD. An interactive chart depicting the percent change (from year 1985) in patient management markers for critical care medicine was constructed:

Big Data

There is no area in medicine where more data is collected than in the intensive care unit. However, the biggest challenge is bringing it all together and constructing a high-quality database. The fact, that there is a lack of unity in the technology and systems being used to monitor patients, makes the whole process extremely tedious.

Another problem is, that we’ve yet to accustom to the digital age, which makes tasks like these less expensive, more precise and a lot easier. Studies have shown there is a significant difference in value accuracy between databases with manual data entry and electronic upload of data — up to a 40% increase in accuracy for digitally collected data.

The largest ICU databases include:

• Adult Patient Database: The Australian and New Zealand database is the biggest with over 400 — thousand patients, developed for central and local analysis and comparison of performance in ICUs.
• Danish Intensive Care database*: Obtains data from the Danish National Registry of Patients and includes virtually all ICU admissions in Denmark since 2005 on average — about 30,000 patients per year.
• MIMIC and eICU: Licensed by MIT and supported by the the NIH, they are the largest easily and freely accessible critical care databases. Their open-source nature makes them effortless to use and collaborate with other researchers
• Project IMPACT*: Was developed by the the Society of Critical Care Medicine and used for validation of prognostic models for classification of patients. The database shows good internal validity for most of the abstracted variables, however poor agreement on continuous data.

* not publicly available

Available data

A huge benefit of intensive care is the fact that patients are monitored 24/7. Time-series data from vital monitors is available periodically and a broad spectrum of information about the patient is collected.

Tables in the databases listed above are usually structured in the following manner:

• Admission: Contains information such as time of admission and discharge, type of admission, preliminary diagnoses or age and gender.
• Procedures and diagnoses: Coded into ICD-9 or ICD-10 provide information about the comorbidities of a patient.
• Periodic: Vital information such as heart-rate, respiration and oxygen saturation is collected on a minute to minute basis.
• Aperiodic: Includes hourly and daily lab measurements (creatinine, white blood cell count and hemoglobin levels) and blood gas analyses.

Prediction in the ICU

Mortality prediction models usually have one or more of the following outcome measures:

1. In-hospital mortality prediction
2. Hospital/ICU length of stay
3. Duration of hospital mechanical ventilation
4. Risk of needing an active treatment during the ICU stay
5. Potential transfers from the ICU

These play a vital role in benchmarking (process of comparing an ICU station’s performance with an external standard) and internal assessment of care. Predominantly, models with a high predictive power could optimise the management of ICU patients and excessively reduce hospital costs.

The main prognostic models for assessing the overall severity of illness in critically ill adults are Acute Physiology and Chronic Health Evaluation (APACHE), Simplified Acute Physiology score (SAPS) and Mortality Probability Model (MPM).

To this day, no model predicting patient outcome has managed to excel in performance over longer periods of time and on a global scale.

Most alarming are the sensitivity scores (correct classification of deceased patients) of the current models. Therefore, further research is necessary in this area.