Crystal Bone Algorithm Predicts Early Fractures, Uses ICD Codes Crystal Bone Algorithm Predicts Early Fractures, Uses ICD Codes

The Crystal Bone (Amgen) novel algorithm predicted 2-year risk of osteoporotic fractures in a large dataset with an accuracy that was consistent with FRAX 10-year risk predictions, researchers report.  

The algorithm was built using machine learning and artificial intelligence to predict fracture risk based on International Classification of Diseases (ICD) codes, as described in an article published in the Journal of Medical Internet Research.

The current validation study was presented September 9 as a poster at the American Society of Bone and Mineral Research (ASBMR) 2022 Annual Meeting.

The scientists validated the algorithm in more than 100,000 patients aged 50 and older (ie, at risk of fracture) who were part of the Reliant Medical Group dataset (a subset of Optum Care).

Importantly, the algorithm predicted increased fracture in many patients who did not have a diagnosis of osteoporosis.

The next steps are validation in other datasets to support the generalizability of Crystal Bone across US healthcare systems, Elinor Mody, MD, Reliant Medical Group, and colleagues report.

“Implementation research, in which patients identified by Crystal Bone undergo a bone health assessment and receive ongoing management will help inform the clinical utility of this novel algorithm,” they conclude.

At the poster session, Tina Kelley, Optum Life Sciences, explained: “It’s a screening tool that says: ‘These are your patients that maybe you should spend a little extra time with, ask a few extra questions.'”

However, further study is needed before used in clinical practice, she emphasized to Medscape Medical News.

A Very Useful Advance but Needs Further Validation

Invited to comment, Peter R. Ebeling, MD, outgoing president of the ASBMR, noted that “many clinicians now use FRAX to calculate absolute fracture risk and select patients who should initiate anti-osteoporosis drugs.”

With FRAX, clinicians input a patient’s age, sex, weight, height, previous fracture, [history of] parent with fractured hip, current smoking status, glucocorticoids, rheumatoid arthritis, secondary osteoporosis, alcohol (≥ 3 units/day), and bone mineral density (BMD) (by DXA at the femoral neck) into the tool, to obtain a 10-year probability of fracture.

“Crystal Bone takes a different approach,” Ebeling, from Monash University, Melbourne, Victoria, Australia, who was not involved with the research, but who disclosed receiving funding from Amgen, told Medscape Medical News in an email.

The algorithm uses electronic health records (EHRs) to identify patients who are likely to have a fracture within the next 2 years, he explained, based on diagnoses and medications associated with osteoporosis and fractures. These include ICD-10 codes for fractures at various sites and secondary causes of osteoporosis (such as rheumatoid and other inflammatory arthritis, chronic obstructive pulmonary disease, asthma, celiac disease, and inflammatory bowel disease).

“This is a very useful advance,” Ebeling summarized, “in that it would alert the clinician to patients in their practice who have a high fracture risk and need to be investigated for osteoporosis and initiated on treatment. Otherwise, the patients would be missed, as currently often occurs.”

“It would need to be adaptable to other [EMR] systems and to be validated in a large separate population to be ready to enter clinical practice,” he said, “but these data look very promising with a good [positive predictive value (PPV)].”

Similarly, Juliet Compston, MD, said: “It provides a novel, fully automated approach to population-based screening for osteoporosis using EHRs to identify people at high imminent risk of fracture.”

Compston, emeritus professor of bone medicine, University of Cambridge, UK, who was not involved with the research but who also disclosed being a consultant for Amgen, selected the study as one of the top clinical science highlights abstracts at the meeting.

“The algorithm looks at ICD codes for previous history of fracture, medications that have adverse effects on bone — for example glucocorticoids, aromatase inhibitors, and anti-androgens — as well as chronic diseases that increase the risk of fracture,” she explained.

“FRAX is the most commonly used tool to estimate fracture probability in clinical practice and to guide treatment decisions,” she noted. However, “currently it requires human input of data into the FRAX website and is generally only performed on individuals who are selected on the basis of clinical risk factors.”

“The Crystal Bone algorithm offers the potential for fully automated population-based screening in older adults to identify those at high risk of fracture, for whom effective therapies are available to reduce fracture risk,” she summarized.

“It needs further validation,” she noted, “and implementation into clinical practice requires the availability of high-quality EHRs.”

Algorithm Validated in 106,328 Patients Aged 50 and Older

Despite guidelines that recommend screening for osteoporosis in women aged 65 and older, men older than 70, and adults aged 50-79 with risk factors, real-world data suggest such screening is low, the researchers note.

The current validation study identified 106,328 patients aged 50 and older who had at least 2 years of consecutive medical history with the Reliant Medical Group from December 2014 to November 2020 as well as at least two EHR codes.

The accuracy of predicting a fracture within 2 years, expressed as area under the receiver operating characteristic (AUROC), was 0.77, where 1 is perfect, 0.5 is no better than random selection, 0.7 to 0.8 is acceptable, and 0.8 to 0.9 indicates excellent predictive accuracy.

In the entire Optum Reliant population older than 50, the risk of fracture within 2 years was 1.95%.

The algorithm identified four groups with a greater risk: 19,100 patients had a threefold higher risk of fracture within 2 years, 9246 patients had a fourfold higher risk, 3533 patients had a sevenfold higher risk, and 1735 patients had a ninefold higher risk.

Many of these patients had no prior diagnosis of osteoporosis

For example, in the 19,100 patients with a threefold greater risk of fracture in 2 years, 69% of patients had not been diagnosed with osteoporosis (49% of them had no history of fracture and 20% did have a history of fracture).

The algorithm had a positive predictive value of 6% to 18%, a negative predictive value of 98% to 99%, a specificity of 81% to 98%, and a sensitivity of 18% to 59%, for the four groups.

The study was funded by Amgen. Mody and another author are Reliant Medical Group employees. Kelley and another author are Optum Life Sciences employees. One author is an employee at Landing AI. Two authors are Amgen employees and own Amgen stock. Ebeling has disclosed receiving research funding from Amgen, Sanofi, and Alexion, and his institution has received honoraria from Amgen and Kyowa Kirin. Compston has disclosed receiving speaking and consultancy fees from Amgen and UCB.

ASBMR 2022 Annual Meeting. Presented September 9, 2022. Abstract FRI-471.  

For more diabetes and endocrinology news, follow us on Twitter and Facebook.