Evaluating postoperative patients in their natural functional environments is an important extension of measuring patient outcomes. This is most commonly done in orthopedics with validated function and quality of life surveys based on subjective patient recall and perception of their well-being (Patient Reported Outcomes: PROs such as WOMAC, VR-12 etc).
The value of these surveys is well accepted when one is reviewing aggregate data sets looking at and comparing large populations, but PROs have not been shown to be uniformly accurate or useful when applied to the measuring the progress of an individual patient. This is primarily because that is not what PROs were designed to do. PROs were intended to quantify the overall change of state from before an intervention to the time of maximum improvement following that intervention. They are not designed to track patients in between those the two endpoints of measure. Further, they generally have ceiling and basement effects that limit their ability to capture the breadth of potential outcomes.
In arthroplasty (i.e.: total hip and knee replacements), maximum improvement is generally considered to be at approximately one year following the operation, leaving the clinician with few validated tools with which to evaluate a patient’s progress during their recovery period. Despite this inherent limitation in the PRO, and because of the lack of agreed upon alternatives, these outcome measures are commonly used to evaluate patients at 6 weeks, 3 months and 6 months.
PROs are therefore the de facto, if not particularly accurate, gold standard of measure for evaluating both a patient’s baseline pre-operative function and their clinical results at any point of clinical recovery following total joint arthroplasty and many other procedures.
The rapid evolution and accessibility of activity monitors may change this unfortunate status quo by giving the healthcare practitioner access to objective functional data as a separate measure to the patient’s own activity recall. The ability to assess out-of-office function on a real-time basis and to evaluate the rate of their recovery is, therefore, a conceptually achievable aim through data collected by smart activity monitors. Further, through blue tooth or direct connectivity to smartphones and computers, activity monitors can communicate automatically with online databases, potentially allowing for real-time tracking and evaluation of accelerometer and gyroscope-based data against future normative values. This, in turn, would create the opportunity for counseling patients on how to optimize their return to function either through the app to the clinician or through feedback directly to patients.
Consumer grade sensors are being widely marketed to (and adopted by) consumers for athletic and wellness purposes but are still in their infancy with respect to applications in the clinical space. Indeed, it is unclear if the data from existing commercial grade activity sensors is sufficiently accurate and reproducible to provide meaningful insight into clinical outcomes.
Furthermore, we do not currently have the tools necessary or the scientific norms in place for how to evaluate and report data thus generated to clinicians and caregivers. It’s relatively new territory. The sheer quantity of data points collected by sensors far exceeds what we have traditionally addressed. Sensors that collect five independent values every second and generate separate metrics derived from these values can create a staggering amount of data over weeks, months and years of use for each individual. In healthcare, we have not in the past had access to such “Big Data” sets, particularly not in the context of research studies or individual patient care.
But we are totally alone in this regard: everyone else it seems has figured it out. Companies such as Google, Amazon, all our banks and social media platforms use software applications such as Hadoop to ingest massive structured and unstructured data sets to extracting meaning and direct decision making. Artificial intelligence algorithms taking advantage of the increasing computational power of modern computers and updated machine learning software can discern complex patterns and associations within the data. It is therefore possible that combining these new tools, the data acquisition platforms inherent in the wearable sensor technologies and the insight available through machine learning algorithms applied to massive data sets can provide clinicians and patients the missing link in our care model: a way to accurately predict future outcomes based on real-time collection of data representative of clinical function.
Investigating the relationship between large amounts of objective patient activity data generated by consumer-grade wearable sensors and subjective validated patient-reported outcome measures (HOOS, KOOS, VR12) was what we set out to evaluate using artificial intelligence in the past year at UCSF. In our pilot study of 25 patients undergoing joint replacements, we collected over 340,000 data points to which we added the clinical record and radiographic reports and PRO data.
Engaging Cloudmedx as our analytics platform allowed us to derive relationships and incredible insights using state of the art ML algorithms. Rather than the standard and rather useless calculation of relative Hazard and Risk Ratios of one outcome versus another when looking at one variable and adjusting for all others (“useless” because the resulting data is incredibly hard to apply in clinical practice where patients present with multiple variables many of which like gender and age are not at all variable), the algorithm was able to clearly identify cohorts (clusters) of people whose variables (features) were more likely to be associated with a given outcome (PRO) and take any specific candidate and place them in a risk cluster.
This is phenomenal and way more practical: give me any unique patient and their data and we can ‘cluster’ them for you pretty accurately.
Very preliminary data was presented this spring at the Orthopaedic Research Society and, as our analysis continues, we will share more. Suffice it to say that we learned a great deal about what kind of data is predictive of what kind of outcome and that we can predict 6-week PROs with a high degree of accuracy in a given individual (R-squared values around 0.7) as early as 2 weeks following surgery. It is very exciting data, or better, “signal” that promises to open an exciting new field of research and allow us to move away from population-based outcomes measures to Precision Medicine style personalized outcome predictive analytics. Welcome, new world. Not a moment too soon.
Dr. Bini is the Founder and Chair of the Digital Orthopaedics Conference San Francisco and a Professor in the Department of Orthopaedic Surgery at the University of California San Francisco (UCSF). His areas of interest include the impact of digital medicine on orthopedic care delivery, change management strategies in health care, and improving the results of total joint surgery. When not working, he is likely spending time with his family or learning to play the guitar. Find him on Twitter @sbinimd.