Patient outcome prediction has been a challenge long reserved for the perceptive eye of the experienced healthcare provider. Today, computers equipped with robust algorithms are learning to extrapolate — and with great accuracy — where a patient’s health is headed. This is an extraordinary benefit in medicine because accurate predictions can facilitate earlier interventions and promote lifesaving preventative care. Doctors are then able to use these predictions to make highly informed decisions regarding a patient’s care.

The prediction process itself is about gathering and analyzing the data that’s already being collected by a hospital’s medical computers. This includes Electronic Health Records (EHRs), imaging, demographic data, and other clinical and administrative health information. Machine Learning is then applied to these stores of information to draw conclusions about potential patient outcomes and to help providers determine possible treatment methods. 

What would be an enormous amount of data for a doctor to manually review becomes a commonplace task for a computer, and the two work in tandem to each do what they do best. Computers scour the data for red-flags and healthcare providers contextualize the information to make the best treatment decision.

Machine Learning in Healthcare

Machine learning is an application of artificial intelligence (AI) that uses algorithms to mimic the way the human brain works and learns. While it can collect various inputs of data and output new sets of data (predictions), machine learning is still more dependent on human interaction in order to learn effectively. Examples include recommendation engines on your favorite movie-watching apps, self-driving cars, and — as we will explore here — patient outcome predictions. Deep learning, on the other hand, is a subfield of machine learning and can function more autonomously. It can make sense of unstructured raw data and categorize the information on its own. Deep learning is more often seen in computer vision technology, face recognition, and speech recognition software. 

Computational Systems Help Determine Treatment Methods

One of the functions of machine learning in this context is for the analysis of large quantities and various forms of patient information. The idea is that the algorithm will find trends among patients with the same diagnosis and make predictions according to how each patient’s data aligns with other cases that have a similar case progression. With this knowledge available, doctors can determine how a patient’s condition may change over time and make care recommendations that take this into account.

Predicting Whether AML Patients will Relapse

An example of this is the computational system developed to predict whether Acute Myeloid Leukemia (AML) patients will go into remission and which will relapse. The computer was “trained” using information from AML patient histories, bone marrow data, and blood data from healthy individuals. 

The complicated part was that the computer would have to take this information and extract further knowledge from it. It wasn’t just going to reorganize it; it would have to make predictions about future data that it hasn’t yet seen. This is the crux of what machine learning is all about. 

In the end, the system was able to predict remission with 100% accuracy and correctly predict relapse in 90% of relevant cases. Having such knowledge in advance allows doctors to take preventative treatment measures for at-risk patients and increase the odds of catching relapse in its earliest stages.

Computational systems like this one are often compatible with a hospital’s existing medical panel PCs. Conversely, the model developed by Cleveland Clinic Cancer Center researchers and their colleagues is still fairly new. As a model with great precision and which offers insight into the biology of AML, the hope is to introduce it to the clinical setting over the next few years.

Determining Progression of Prostate Cancer

Even with the ability to predict accurately, knowing there is a cancerous presence in the body does not inform doctors about how aggressive the malignancy is or how well it will respond to certain therapies. 

For conditions like prostate cancer, understanding the severity of the condition can be hugely beneficial. Some patients with the disease may never need treatment while others need the prostate fully removed. Without the ability to predict how an individual patient’s tumor is going to progress, some men may not get the aggressive therapy they need while others may get a radical therapy they don’t need.

This is where predictive computational models come in. Taking into account variables like a patient’s age, the estimated extent of the disease, the estimated aggressiveness of the cancer, comorbidity, and other factors, researchers developed a model that could generate a prognosis for patients on an individualized basis. This particular model predicted death due to prostate cancer with 85% accuracy, showcasing its great potential to reduce both over- and undertreatment of the disease.

Similar to how the AML system functions, this computational model, termed PREDICT Prostate, assesses the risk associated with each disease variable to come up with its prognostic prediction.

Predicting Cardiovascular Collapse

When humans suffer blood loss, our bodies have a number of automatic responses designed to compensate for this loss. A number of undetectable mechanisms begin to take place — like vasoconstriction, improved cardiac filling, and more efficient breathing. By the time vital sign metrics indicate that something is amiss, as much as a third of circulating blood volume may have already been lost. By that point, the body has been in a state of stress long enough that it may soon begin to run out of steam to keep those compensation responses going. When this happens, the patient could enter cardiovascular collapse.  

Flashback Technologies developed an algorithm called the Compensatory Reserve Index (CRI) to help predict when a patient was about to go into cardiovascular collapse. The system monitors the patient and learns their individual compensatory mechanisms. Those normally undetectable responses are then readable as arterial waveforms. It can detect when those mechanisms change and, as a result, predict when their body is about to decompensate. 

Initially devised to help soldiers who were wounded in the field, the system is now used by first responders in civilian settings to detect hemorrhagic shock more quickly. 

Reducing Length of Stay and Readmission

Predictive modeling is achieving impressive results in helping gauge the course of a person’s disease and it’s also playing a role in identifying patients who are at risk for an adverse event. These are patients who face readmission to the hospital, an extension to their length of stay (LOS), and who are likely to die in the hospital. In this use-case, the intent of using these computer findings is to better understand the various risk factors and their interactions. 

There is a complexity behind these adverse patient experiences that limits the ability of machine learning to predict the probability of readmission and extensions to length of stay. However, key variable interactions are visible when the data is investigated. For example, high and low heart rate values at discharge were shown to affect the probability of readmission. Older patients with lower heart rates upon discharge were less likely to be readmitted, while higher heart rates at discharge meant they would be more likely to be readmitted. Pinpointing information like this allows for more nuanced research in how machine learning can contribute to patient care.

In other scenarios, algorithms are predicting whether COVID-19 patients will need ICU intervention. Through the analysis of CT images, demographic information, vital signs data, and lab blood test results, machine learning is helping doctors digest large sums of data to determine what kind of treatment the patient will end up needing. So far, success with this system indicates that other conditions like lung disease can benefit from similar computational applications.

Humans Make Up for AIs Limitations

Artificial intelligence is picking up steam in the realm of healthcare but that’s not to say that clinical judgment is rendered obsolete. Machines do what they do best; — and admittedly they do so better than humans — they process, analyze, and find patterns in data quickly and without bias. Healthcare providers, however, have the distinct advantage of being able to interact with patients face-to-face and to glean information that is often missing from patient records. 

Key psychological and emotional details, even if documented in plain text, are usually not recorded in a way that allows algorithms to successfully utilize the information. This is why human intuition is still a critical component of effectively implementing machine learning in healthcare. 

The other fundamental detail is the way that AI interacts with medical computers in a healthcare environment. Understanding how shifts in technology impact existing computer systems determines how prepared a facility is to take advantage of emerging trends. If your organization is considering AI solutions, contact our team to discuss which of our medical grade PCs are best suited for your needs.