There’s a saying: “The best way to predict the future is to study the past.” While predicting the future has often seemed a distant concept, reserved for the realm of science fiction, what if in healthcare there are ways to use past data? By studying patterns in medical data, we can gain valuable insights into a patient's future health outcomes.
How? By leveraging Electronic Health Records (EHR) and predictive analytics, healthcare organizations can foresee potential risks. For example, a predictive analytics model might highlight a patient with heart failure who has a higher probability of being readmitted within 30 days due to multiple past hospitalizations and non-compliance with prescribed medications.
Our Head of AI research, Sandro, was already talking about EHR and how his team helped clients implement FHIR in their software solutions, and for this blog post , I went through several studies and focused on three key findings. Let’s explore where predictive analytics can help us forecast the future and, most importantly, how it can benefit patients.
1. Readmission risk
I recently read research in which they developed a deep learning framework to predict 30-day unscheduled hospital readmissions for Congestive Heart Failure (CHF) patients. They created a model using real Electronic Health Record (EHR) data from over 7,500 CHF patients in Sweden hospitalized between 2012 and 2016. The goal of this model was to improve the identification of high-risk patients at the time of discharge.
In this context, "high-risk patients at the time of discharge" refers to patients who are more likely to be readmitted to the hospital within 30 days after they leave. These patients typically have certain characteristics or medical conditions that make them more vulnerable to complications after they are discharged. This all leads to unscheduled readmissions.
The goal of the study was to create a model that can accurately predict which patients are at this "high risk" when they are discharged. Identifying these patients in advance allows healthcare providers to take special steps (targeted interventions) to help prevent their readmission.
One of the findings was that including sequential patterns (the patient's history of clinical visits) improved the prediction accuracy by 26%.
I highly recommend reading the article. It shows that one of the most valuable applications of predictive analytics in healthcare is its ability to forecast hospital readmission risk.
Hospital readmissions can be costly, but even more important is they can negatively impact patient outcomes.
Using electronic health records (EHR) can help healthcare providers anticipate which patients are at higher risk of being readmitted after a hospital stay.
Hospitals and healthcare systems can analyze historical healthcare data, such as past treatments, patient demographics, and previous hospitalization patterns, to identify trends and risk factors.
2. Disease progression
Predictive analytics is also a powerful tool in understanding and managing the progression of diseases, particularly chronic conditions.
By analyzing historical data stored in EHR systems, healthcare professionals can gain insights into how a patient's disease has progressed over time. This allows for more personalized care and early intervention. This is especially beneficial for diseases that are progressive and chronic, that if not properly managed can turn much worse.
In this research, you can read how researchers developed a recurrent neural network (RNN) model to predict the progression of chronic kidney disease (CKD) from early stages to more severe stages. To do this, they processed longitudinal electronic health records (EHRs), which include data such as lab tests, demographic details, vital signs, and health behaviors.
Considering CKD is a progressive disease, it can require dialysis or a transplant. It seems that using longitudinal data (which refers to the data that is collected over time, allowing for observation of changes, trends, patterns, etc) from EHR improved the prediction of disease.
This helps prevent the worsening of chronic conditions and enables patients to maintain a better quality of life. The combination of predictive analytics solutions and historical healthcare data represents a leap forward in chronic disease management.
3. Early detection of disease
One of the most transformative uses of predictive analytics in healthcare is the early detection of disease. By leveraging vast amounts of patient data from electronic health records, healthcare providers can now anticipate the onset of diseases before they manifest with noticeable symptoms.
When we talk about predictive analytics in healthcare, we often refer to the structured data from EHRs, but recently, I read research that incorporated both structured and unstructured data from EHRs can improve chronic disease prediction.
In the article "Deep EHR: Chronic Disease Prediction Using Medical Notes," they discuss the development of a machine learning model that uses both structured and unstructured data from Electronic Health Records (EHR) to predict the onset of chronic diseases.
The study aimed to improve chronic disease prediction by utilizing both structured data (like lab results and demographics) and unstructured free-text medical notes. The motivation for the study is based on the observation that a vast amount of valuable clinical information is contained in free-text medical notes, which are often neglected in predictive models. Most prior disease prediction models have focused solely on structured data such as lab results and diagnosis codes, but these structured fields miss out on a lot of the nuanced information present in narrative notes.
The findings were that models using unstructured medical notes consistently outperformed models that used only structured data (like lab values and demographics). This demonstrates that text data contains important predictive information that structured data alone misses.
What about ethical considerations?
The findings I read showcase that the use of predictive analytics in healthcare offers actionable insights that enable healthcare organizations and medical professionals to develop proactive care plans.
For instance, by analyzing patient data, healthcare providers can predict which patients have a high risk of readmission and arrange follow-up appointments or additional support to prevent unnecessary hospitalizations.
However, several challenges must be addressed. The effectiveness of predictive analytics models depends heavily on the accuracy of the data they analyze. We've mentioned this in several blogs, and it holds true whether in healthcare or software development: the quality of the data you input directly impacts the quality of the results you get.
Incomplete or inaccurate EHR data can lead to incorrect predictions, potentially compromising patient care. Additionally, as more data is collected and analyzed, patient privacy becomes a concern. Although regulations like HIPAA aim to protect patient information, breaches can still occur, posing risks to sensitive health data. Moreover, if the data used to train AI models is biased or unrepresentative, the resulting predictions may not be correct.
While AI has recently become a buzzword, it's important to remember that this technology has been developing in the background for many years. Predictive analytics in healthcare—though there are challenges, such as incomplete data and concerns about patient care, we are still in the early stages of its potential. I believe this is just the beginning. We need to be cautious about these limitations, but it’s also a reminder that AI is simply a tool designed to assist humans, not replace them. Ultimately, healthcare will always require the human touch, with medical professionals and AI working together to enhance patient care and outcomes.
I’ll leave this here: in the article I referenced earlier, it was found that adding expert (human-derived) features boosted performance by 3%. Expert human-derived features refer to features that are manually selected or designed by domain experts (such as doctors or healthcare professionals) based on their understanding of the data and the problem at hand.
