Blog from July, 2019

Samara Rosenfeld

Inpatient violence risk assessments can be performed automatically using available clinical notes without sacrificing predictive validity, according to the findings of a study published in JAMA Network Open.
Researchers used machine learning to analyze clinical notes in the electronic health records (EHRs) of two psychiatric institutions to predict inpatient violence. Investigators measured each site’s area under the curve to determine predictive validity. The first site had an area under the curve of 0.797, while the second registered at 0.764, meaning it is possible to use routinely registered clinical notes for automatic violence risk assessment. 

The model performed with a specificity between 0.935 and 0.947 and a sensitivity between 0.334 and 0.336.
“Inpatient violence remains a significant problem despite existing risk assessment methods,” the study authors wrote. “The lack of robustness and the high degree of effort needed to use current methods might be mitigated by using routinely registered clinical notes.”
The research team used the first site, the psychiatry department of the academic medical center in Utrecht, Netherlands, for internal method validation. The data set consisted of 3,201 admissions of 2,211 unique patients. The second site, a general psychiatric hospital that delivers secondary care in Rotterdam, Netherlands, was used for external method validation. This data set consisted of 3,277 admissions of 1,937 unique patients.
Researchers extracted clinical notes written by psychiatrists and nurses from patients’ EHRs. The research team included notes written in the four weeks before admission up to the first 24 hours of admission. They excluded admissions with fewer than 100 words registered after 24 hours.
Reports of violent incidents helped determine the outcome for each admission. Staff members involved in a given incident at either site filled in structured information, a textual description of the incident and the severity of the incident as measured by the Staff Observation Aggression Scale-Revised.
Violent incidents included all threatening and violent behavior of verbal or physical nature directed at another person. This excluded self-harm and inappropriate behavior like substance abuse or vandalism.
Researchers examined predictive power hidden in the notes by extracting the 1,000 most frequent terms. A chi-squared test helped assess the strength of the term’s association with the outcome. The research team selected the top 10% of predictors based on their chi-squared test scores in 1,000 repeated samples.


Terms such as “aggressive,” “angry,” “verbal,” “threatening” and “irritated” can be directly associated with violence. Terms like “reacts,” “walks” and “speaks” describe behavioral cues that can be indirectly associated with violence.
Researchers then used a machine-learning approach to perform a violence risk assessment. Algorithms can detect patterns in historical data, and prediction can help the course of treatment based on those patterns.
The approach transformed clinical notes into a numerical representation and then fed the representations into a classification model.
Researchers trained the model using the internal set of clinical notes.
The risk of violent outcomes for patients with predicted high risk compared to low risk was 5.121 in the first site and 6.297 in the second.
“In the near future, we envision that further advancements towards a data-driven psychiatric practice will be made and that EHR data will become an even more valuable asset in supporting important decisions in the clinical practice,” the authors wrote.

Nathan Eddy

Pre-treatment scans were input into a deep-learning model, which analyzed the scans to create an image signature that predicts treatment outcomes.

Artificial intelligence and machine learning networks could help personalize radiation therapy for lung cancer, according to a new study by the Cleveland Clinic.

The research, published in The Lancet Digital Health, centers around an artificial neural network built with a large dataset of patients receiving lung radiotherapy.

That network, which allows each clinical center to utilize their own CT datasets to customize the framework and tailor it to their specific patient population, was built using CT scans and the electronic health records of nearly a thousand lung cancer patients treated with high-dose radiation.

The company's framework uses probability estimates to select an optimized dose that prevents treatments failures to a set level, for instance a five percent probability of failure.

Pre-treatment scans were input into a deep-learning model, which analyzed the scans to create an image signature that predicts treatment outcomes.

This image signature was combined with data from patient health records, to generate a personalized radiation dose using advanced mathematical modeling.

"AI can learn from imaging and electronic health records and make predictions about the likelihood an individual patient could fail radiation treatments," lead author Dr. Mohamed Abazeed, a radiation oncologist at Cleveland Clinic's Taussig Cancer Institute and a researcher at the Lerner Research Institute, told HealthcareITNews.

"Therefore," he said, "AI can help individualize radiotherapy treatments for patients with cancers in the lung."

Dr. Abazeed explained they will assess the transportability of the model across varied hospital systems via local implementation or using large-scale federated datasets.

In the future AI-models could be optimized based on different target populations based on ethnicities, gender or age, medical settings (community hospital or academic center) geographical locales  or even include temporally distinct populations.

"We will also test the putative supremacy of iGray--individualized dose--recommendations head-to-head with standard of care recommendations in a prospective clinical trial," Dr. Abazeed said.

In reference to those who believe AI technology still has much farther to go before it has practical applications for the medical and healthcare sectors Dr. Abazeed noted a prerequisite for scientific progress is the willful suspension of disbelief.

"In large part driven by this work, we are on the precipice for practical and innovative implementations in the highly standardized and data-replete discipline of radiation oncology," he said. 

The study follows news that French biopharmaceutical company Sanofi and tech giant Google are partnering to leverage machine learning, AI and deep analytics technologies across data sets to better understand major diseases.

Meanwhile, a new study from Innovaccer explores the ways its AI algorithms could be put to work to improve risk scoring and stratification and enhance value-based care initiatives.

Benjamin Harris

Many illnesses are easy to treat if caught in time. New EHR studies show how optimizing data can spot problems clinicians might overlook.

Early detection is a mantra in hospitals: Find a nascent condition before it can morph into something serious and you can alleviate suffering, improve care and save lives.

Symptoms can be hidden in plain sight or can be masked by other known illnesses. In these manners, thousands of patients a year fall through the cracks and become seriously ill.


Two new early detection algorithms that integrate with electronic health record data are looking to make the odds more favorable.

Johns Hopkins has developed an algorithm called TREWS that can detect sepsis in patients far more reliably than a clinician alone, and the American Medical Association has published guidelines on optimizing EHRs to identify patients at risk of diabetes.


In the case of both type 2 diabetes and sepsis, the at-risk population is broad and the conditions for both diseases are widespread: Accurately targeting those most likely to be affected takes more detailed hunting than many clinicians can do.

Johns Hopkins researchers note that one in 10 sepsis alerts from a “dumb” EHR system are true – their algorithm brings that down to one in two. Similarly, the AMA says that while many Americans are at risk of type 2 diabetes, optimizing an EHR to monitor relevant test results and make it easy for clinicians to order additional testing – both steps that empower providers to better manage their population health.

“One of the challenges traditionally with sepsis has been making sure the patient gets all of the interventions within the first three hours,” says computer science graduate student Katharine Henry, who worked on the Johns Hopkins TREWS program.


Artificial intelligence-assisted detection technologies are helping clinicians find needles in haystacks in every specialty.

When a disease like sepsis can present through very common symptoms like elevated heart rate or temperature, a doctor can’t be sure of a false positive based on a few metrics. Relying on an optimized EHR algorithm that monitors more data points than a clinician alone could, however, means the EHR does the heavy lifting on the data end and frees a provider up to take action when needed.

Similarly, much like a disease like type 2 diabetes can be brought under control through management of lifestyle, a healthcare provider can nip many instances in the bud when they manage the health of their patient population well. Providing monitoring tools and predefined courses of action through an EHR makes identifying patients and taking appropriate action that much easier.