The Problems With Early Warning Scores and How We Can Fix Them

Written by Dr. Dinesh Rai, MD

Early warning scores (EWS) help medical staff identify when a patient’s condition is deteriorating. They trigger an alert to indicate a patient who should be monitored more closely and, if necessary, given more intensive treatment. There are many different early warning scores in use, but they all work broadly the same way. The patient’s vital signs (i.e., temperature, heart rate, respiratory rate, and blood pressure) are measured and assigned a numerical score. Some scores add additional features like level of alertness, oxygen delivery methods, and bloodwork. These scores are then combined and weighted to give an overall score to assess the risk of patient deterioration and recommend potential assessment and frequency of monitoring. Early warning scores are valuable because they provide the healthcare team with an objective, standard way to assess the patient’s risk of deterioration. These scores have been shown to correlate with increased risks of mortality. The Modified Early Warning System (MEWS) was associated with a roughly 50% decline in non-ICU code blues. Provider judgment is still superior. One study used a “Dutch-early-nurse-worry-indicator-score” to alert clinical teams and initiate rapid responses before EWS are triggered with superior outcomes. Although helpful in adding color to a patient’s clinical picture, EWSs have room for improvement.


Problem #1: Lack of Automation

Early warning systems require vital sign information to be somehow transferred to them. These often take the form of nurse input, which can be time-consuming for staff. Manual entry could lead to human error (incorrect inputs, an incorrect tool used) and cause delays in reporting such scores. 

Opportunity: Automation of EWSs could reduce the chances of human error and eliminate delays in reporting scores to nurses and health care providers. A system that can ingest vital sign data from monitors throughout a healthcare system and push timely notifications to appropriate staff can improve the utility of these systems.


Problem #2: Lack of Specificity

One of the biggest problems with EWS is that they tend to have low specificity and a predilection towards false alarms. EWS can have up to 77% false positive ratio. These false alarms impact both staff and patients negatively. Alarm fatigue is a term used to describe the effect of continually receiving false alarms. Alarm fatigue leads staff to ignore or turn off alarms, causing delayed responses and missed interventions. The consequences of this can be fatal if other alarms for severely deranged vitals are ignored. EWS may trigger staff to frequently evaluate patients and lead to unnecessary testing, causing a drain on resources. Patients are subjected to increased intrusions, examinations, and awakenings.


Problem #3: Lack of Temporality

Rapid changes in vital signs, even if they remain in the normal range, could indicate clinical status changes. A patient’s heart rate increasing from 60 to 90 (both within the ‘normal range’) over the course of hours may indicate deterioration and be a good indicator for early warning. However, these EWS systems don’t take temporality into account. The vital sign inputs are a snapshot of the patient’s clinical status. This is akin to looking at a still and trying to understand the movie’s story. A picture can give you an idea of what is going on, but including temporality and trajectories will provide you with a better idea of where a patient is headed physiologically.

Opportunity: Zhu et al., out of the University of Cambridge, developed a dynamic EWS called dyniEWS. DyniEWS assesses a patient’s snapshot vitals and other statistical analyses, including deviations from the patient’s median, the most recent rate of change, and a host of other features. The most critical features of DyniEWS were determined to be a snapshot of oxygen concentration (FiO2), patient consciousness level, an average of all sequential FiO2 measurements, and frequency of measurements. Note that two of these features cannot be determined from a snapshot picture. Machine learning can play an essential role in introducing more temporality into EWSs. Slight changes in vital signs and their associations with other factors could predict deterioration well before it happens. However, these changes may be too numerous for a traditional rule-based system. A deep learning model can be trained on large volumes of vital signs time series data to reveal subtle associations that are difficult for humans to uncover.


Problem #4: Lack of Flexibility

EWSs do not consider patient-specific factors, including medical history, primary chief complaint, and baseline physiology. Patients with specific diseases have baseline vital measurements outside the normal range. Chronic obstructive pulmonary disease patients often have baseline FiO2 measurements of 88-92%. End-stage renal disease patients tend to have high baseline blood pressure. Normal-range vital signs in these patients could be of clinical concern. There are also disease-specific vital sign parameters that are outside normal. Trauma patients have systolic blood pressure targets of 80-90mmHg. Conversely, patients with acute ischemic strokes are permitted systolic measurements up to 200mmHg. 

Opportunity: Targeted EWS parameters can reduce false alarms and trigger alerts when typical EWSs would not. There are several specialty-specific EWS including obstetrics, colo-rectal surgery, and cardiac. Individually adjusted EWS, however, is still an exciting area for research. One study showed that adjusting National EWS (from -4 to +6) based on nursing clinical assessment was non-inferior to National EWS alone. There is potential for machine learning applications in this space as well. There is significant variability in each patient case and creating an effective rule-based algorithm is complicated. Machine learning algorithms again can ingest large amounts of data to derive appropriate correlations. 

CalmWave has studied existing scoring systems and reviewing tens of thousands of patients’ data to determine how to improve upon existing methods of early warning systems, leveraging our deep experience in data science and alarm remediation.  It’s clear that best opportunities to improve upon existing methods can only be achieved via advanced analytical methods, aka artificial intelligence.  The stakes are high in clinical care, and if EWSs are not effective, they have the risk of exacerbating the issue of alert fatigue for providers, and/or hurting patient’s outcomes.



Early warning systems are essential to assist providers in intervening and preventing a patient’s clinical deterioration. They tend to have positive predictive value, can improve outcomes, and help develop a standard way to communicate clinical status. Frequent false positives, lack of automation, patient specificity, and inability to account for temporal changes have hindered their adoption and effectiveness. However, the future of EWSs is exciting and several research developments have shown various ways they can be improved. These improvements will reduce the alarm burden and increase the utility of these alerts.  Request a demo with today to see how we can filter out noise and insights from one of your most valuable assets, your data.