A Way to Detect Major Medical Complications Sooner

A study of surgical patients, across 168 hospitals, showed that 23% of patients experience a major complication during their stay. We like to think of complications as atypical events. However, the unfortunate truth is that they are quite common. While most medical complications are easily identified and are treated in a timely manner, not all are recognized soon enough. And delayed intervention means fewer treatment options and poorer outcomes.

My mother, Florence Rothman, was one of these patients whose complications were recognized too late; she died in 2003 in a hospital of avoidable causes. Her deterioration went unnoticed, and my brother and I have spent the last 15 years working to help prevent that next avoidable death.

There is one question that a clinician does not want to have to answer, “Why didn’t we see this patient’s problem sooner?” To deliver better care, doctors and nurses need to fully understand the patient’s current status in order to predict potential problems. Yet, many hospitals in the United States rely on only vital signs as status indicators and do not capitalize on the full complement of available patient information, especially nursing assessments — each nurse’s careful evaluation of his/her patient’s condition that is conveniently recorded in the electronic medical record (EMR). With this data, I believe that it is possible to implement an “unblinking eye”: a 24/7 evaluation of patient status that leverages patient data more completely. In an age of such tremendous technological innovation, health care must step up and change outdated processes to help save lives by integrating and embracing all patient data to identify deterioration sooner.

By making better use of patient data and predictive models we can identify patients who are “smoldering” (in the words of one nurse): Those who may appear fine but have unseen damaging processes occurring internally. Sepsis can be such a process. It can tragically afflict otherwise well patients and is often identified too late, which is why it is the focus of a major worldwide effort to reduce its death toll.

But making a prediction is not enough. For a prediction to affect patient outcomes, it must meet the criteria that I term the prediction trifecta: It must be correct, timely, and provide new information.

Many prediction models currently in use at hospitals today rely on vital signs to satisfy the first criterion, “correct,” to identify an impending crisis. However, while identifying a patient who is deteriorating is easily done with vital signs alone, it is far more difficult for such a system to meet the next criterion, “timely,” and provide a warning when there are still options available to halt the deterioration.

Systems relying on vital signs rarely meet the third criterion, “new”, which is providing information that is not already known to the physicians and nurses. While vital signs are important and valuable, there are three intrinsic shortcomings in focusing on them for early warning:

Vital signs tend to be lagging indicators. The human body is built to maintain equilibrium with basic, vital operating parameters, and we do so by sacrificing functionality. Appetite fades. Digestion shuts down. Fluid builds in the extremities. All of these effects can happen while we maintain un-alarming vitals, so when the vitals fail, and decompensation is seen, it tends to be too late for effective intervention.
Vitals are generally only available to a predictive model when a nurse or a technician enters the data. The nurse is therefore ahead of the model and a model based on vitals rarely ever provides “new” information. Any predictive model based upon vitals alone is therefore unlikely to be either timely or new.
Further confounding vital sign-based models, normal variation in patients that are not in trouble tend to swamp the signal from those few who are, leading to high rates of false positives, alert fatigue, and clinician tune-out.

The goal of achieving this prediction trifecta is not to replicate what we see in frantic, hospital TV dramas with nurses and physicians racing to a patient’s bedside with a “code blue.” That patient has about a 17% chance of going home, if he or she is even revived.

Clinicians need predictions to be meaningful and receive them early in the deterioration process. Continuing with the sepsis example: In the time lag from inception to treatment, it’s critical to administer a bolus of fluids and IV antibiotics. One estimate claims an increase in mortality rate for every hour of delayed treatment. Mortality rates for early-detected sepsis are about 5%, but if it is allowed to progress, mortality rates approach 50%.

Nursing Assessments: The Key to Meaningful Prediction

Fortunately, there is another source of physiological data recorded periodically for every patient in the hospital’s EMR system. Nursing assessments, the structured evaluation of a patient’s physiological systems, can identify a patient’s deterioration from sepsis or other conditions and complications before it’s evident in vital signs or laboratory data. Yet, many current prediction models do not include this information.

Nurses conduct what’s termed a “head-to-toe” assessment on each patient, every day, every shift, in every hospital. It includes, for example, cardiac, respiratory, gastrointestinal, neurological, skin, psychosocial, and musculoskeletal assessments. For each evaluation, a nurse interacts with the patient to conduct and document a structured, hands-on review. If all the underlying factors of that assessment are normal, then the nurse deems it passed or met; if one or more of the factors is viewed as abnormal, then the assessment will be failed, or not met. For example, your skin is an organ. The skin nursing assessment reviews changes in skin texture, continuity, and color. In sepsis cases, skin failure may be an early indicator.

Every nursing assessment requires human, clinical judgment and provides both insight into the patient’s current state and predictive power to help identify patients who are at an elevated risk of an adverse outcome. Effective predictive models must include these leading indicators.

As an attempt to prevent what happened to our mother from happening again, my brother Steven and I developed a tool called the Rothman Index (RI), which includes nursing assessments, that can help provide an “unblinking eye” to support clinicians. Through the RI and the help of nursing protocols, Yale New Haven Health System has reduced in-hospital mortality by 20% to 30%, with special benefits in reducing sepsis mortality. Meaningful prediction, hitting the trifecta, has also helped the organization see a reduction in cost per sepsis case.

The inclusion of nursing data in predictive models makes profound sense: The nurse understands the patient’s condition. If we capture that nursing gestalt, and especially if we can do it electronically, we are on our way to reducing that critical time lag between inception of a possibly life-threatening complication, and action.

All models must be tested for their value in providing not only prediction but also for their value in providing meaningful prediction. It’s a concept that was inspired by one life, but as a standard practice, can put us on the path to saving countless others.