Using Machine Learning to Predict Patient Outcomes

Moby Parsons, MD

During training and throughout our early careers, we were taught that good results come from experience, experience comes from bad results, and nothing ruins good results like follow-up. The problem with experience is that we tend to base our decision-making on both our best and our worst outcomes. Thus, our disaster patients and our superstar patients are the ones that tend to guide our clinical decision more often than the average patients who may define the gold standard for a condition. Ideally, we should be looking at the entire spectrum of results across the bell curve to guide optimal decision-making in a patient-specific fashion. This is particularly true for shoulder arthroplasty where indications, a better understanding of pathoanatomy, preoperative planning tools, improved surgical techniques, and a wider array of implant options have all changed the landscape of how we treat degenerative shoulder conditions.

Machine learning (ML), a data science tool that is a discipline of artificial intelligence, has the capacity to advance patient-specific clinical decision-making in a manner that is much more powerful than either retrospective or prospective outcomes analysis or surgeon experience. ML is currently used in applications we are intimately familiar with including Internet searches (Google), recommendation engines (Netflix, Amazon), chatbots (online retail custom service), and fraud monitoring (credit card companies). ML is also currently being successfully applied to image analysis in radiology to improve cancer detection and is making waves in other fields of medicine where large data sets can reveal unseen patterns.  Unlike conventional statistical methods which are backward-looking, use small data sets, and rely on inference, ML uses large data sets in a forward-looking manner to make predictions. The true power of ML lies in its ability to capture complex and non-linear relationships in large data sets that rule-based analysis, like regression modeling, may fail to capture. ML can then turn information into actionable insights.

Teaming up with KenSci, a data science company located in Seattle, Wash., Exactech has been at the forefront of using ML to better predict outcomes and complications after shoulder arthroplasty. This work is based on Exactech’s clinical database which includes over 11,000 patient visits from 35 centers around the world–all using a standardized data collection tool that records information on demographics, diagnosis, comorbidities, preoperative function, implant information and post-operative function at multiple time points. Predict+ was built on ML algorithms which established a 19-input minimal feature set that was most highly predictive of outcomes and complications after anatomic or reverse shoulder arthroplasty.

Predict+ provides VAS Pain, Global Shoulder Function, and predicted range of motion at multiple time points up to 7 years after surgery. In addition, it provides age and gender-adjusted complication rates as well as ASES and Constant Score predictions if the surgeon chooses to include these in the input data.  Predict+ can also show metrics like the MCID and the SCB, as well as factors driving the predicted outcomes scores up and down. This allows risk factor mitigation to optimize outcomes after shoulder arthroplasty. Future iterations of Predict+ will include imaging information on boney wear patterns, muscle atrophy, cuff integrity and possibly bone density. All of this information, including the current clinical inputs, can help surgeons decide in advance the optimal surgical intervention and the expected patient-specific outcome over time.

The value of Predict+ lies in its ability to provide evidence-based, patient-specific information that can help set patient expectations and goals after shoulder arthroplasty and enhance surgeons’ clinical decision-making in determining the optimal procedure and whether the margin for improvement is sufficient to benefit the patient. Providing this information to patients allows a more informed decision-making process, which in itself can lead to better patient satisfaction as well as set benchmarks for patients to achieve. Combined with GPS planning and navigation, Predict+ adds to the Equinoxe® Active Intelligence platform of tools that is defining the evolution of shoulder arthroplasty as it moves from an implant-driven service line to a solution-driven service line.  Now integrated into the Equinoxe Preoperative Planning Application, Predict+ can be easily integrated into the clinical workflow of practice and can be downloaded by visiting


Moby Parsons, MD, practices at the Knee, Hip and Shoulder Center in Portsmouth, N.H., and is a member of the ExactechGPS® design team. He trained at the University of Pittsburgh where he performed a research fellowship in shoulder surgery; his clinical research on tendon transfers won the Neer Award. Dr. Parsons received fellowship training at the University of Washington and University of New South Wales. He is a founding member of the New England Shoulder and Elbow Surgeons and a regional leader in shoulder surgery, including outpatient shoulder arthroplasty.

To learn more about Predict+™, Exactech’s patient-specific outcome predictor, visit