Compliance Makeover: Developing Data Intelligence

DMEC Staff@Work

When Small Data Informs Big Data: STD Progression to LTD

Developing Data IntelligenceBy Fraser Gaspar, PhD, MPH


We have all seen claims that ended up with unexpectedly poor return-to-work (RTW) outcomes, and wanted to know why. At a recent conference, attendees were challenged to do a “case post-mortem” analysis. My opportunity to do this arose when I joined a research project to understand the factors that influence case progression from short-term disability (STD) to long-term disability (LTD). The pilot project revealed interesting differences between claims with good or poor RTW outcomes.


To understand the factors that influence progression from STD to LTD, the pilot used STD data from a manufacturer with 8,000 employees across North America and Europe. This firm had 512 STD claims over two years, all with a maximum benefit duration of 181 days, in which 4.1% of cases went on to LTD. We randomly selected 10 STD cases that went to LTD and compared those to 10 cases that did not go to LTD, but returned to full duty. Many employers can support such a pilot effort, which is a first step in developing more sophisticated case management systems that can track additional case factors to extract knowledge from data.

The cases in our pilot project were matched by ICD-10-CM diagnosis or, if unavailable, by code category. If further matching was needed, we progressively matched on whether each claim reached maximum benefits, followed by number of leaves since hire, sex, and then age. We did not include progressive or ill-defined illnesses, such as cancer or multiple sclerosis, or symptom-based diagnoses in this analysis.

With the help of expert nurse case managers, we developed a case review form with 54 questions about case characteristics, including demographics, access to care, psychosocial aspects, and medical treatment (see Figure 1 for examples).1

Example 1
Case Review Example Questions
  • Did the employee express/share motivation to return to work?
  • Is there any motivation to not return to work, including extra benefit payment or avoidance of work?
  • Does employee’s geographic location prevent access to healthcare?
  • Was an opiate prescribed at any time?
  • Did other underlying factors contribute to the disability? Specify.

What Did We Learn?

The most interesting finding was that the “underlying factors” of a disability case were the most predictive of whether the case would go to LTD. These underlying factors ranged from common comorbidities (e.g., diabetes, hypertension) to more complex diagnoses such as a history of joint replacement or neuropathy. Of the 10 cases that went to LTD, seven had an underlying factor. Of the 10 cases that did not go to LTD, only three had an underlying factor (Figure 1).

Figure 2
Underlying Factor Went to LTD?
Case 1: Diabetes, acute asthma, hypertension No
Case 2: Job requirements, repetition No
Case 3: Family history No
Case 4: Depression, anxiety/bipolar, history of polysubstance abuse Yes
Case 5: Smoking Yes
Case 6: Hypertension Yes
Case 7: Diabetes, history of joint replacements Yes
Case 8: Advanced age Yes
Case 9: Diabetes, hypertension, neuropathy, vertigo, reactive attachment disorder, pedal edema, urinary frequency, sinus tachycardia Yes
Case 10: Diabetes Yes

We also found evidence that both positive and negative motivation were significant factors in progression to LTD. Among cases that progressed to LTD, only 10% of claimants had positive motivation to RTW, and 50% had motivation to not RTW (such as anticipating retirement). Among claimants that avoided LTD, 50% had positive motivation to RTW, and none had negative motivation. These results make it clear that leave managers should explore how they can reinforce positive motivation and reduce negative motivation.

In another interesting finding, the “disabling diagnosis” was often different from the starting disability diagnosis. In 25% of all cases, the disabling diagnosis ended up being quite different from the initial diagnosis. One individual initially went out for low back pain (LBP), but the disabling diagnosis was depression. In another case, an individual was on disability for Type II diabetes, but the disabling diagnosis was atrial fibrillation.

It is not difficult to see how these cases could progress. The link between disability and starting diagnosis makes sense as low back disorders are often correlated with depression2 and diabetes is a risk factor for atrial fibrillation.3 Nonetheless, the results highlighted the fact that claims often evolve through time. Careful follow-up and attention by nurse case managers are needed to stay on top of the diagnoses that may prevent a successful RTW.

Beyond looking at LTD as an outcome, when we compared the durations of those with successful RTW to MDGuidelines’ benchmark durations, individuals who RTW were on average out on STD leave for 24 days longer than the benchmark. This number is artificially high due to using non-LTD cases that reached maximum benefit duration during our matching procedure, but is far fewer than the 126 extra days on STD leave observed in the group that went to LTD.

Significance of Findings

This pilot project underscored the complexity of many disability claims. From underlying factors to disabling diagnoses that change over time, a disability claim does not always follow a straightforward path. The challenge of trying to understand complex problems with limited case data can inspire professionals to track more data in the future. Recently, another firm requested a custom analytics job to understand the factors that influenced their disability durations. The only data they could provide were age, sex, industry, and the initial diagnosis (which is often merely a symptom of the underlying condition). It would certainly be convenient if such limited data could explain the variation that occurs in RTW outcome, but this merely represents a starting point.

To capture and use data to improve outcomes, I suggest two paths.

First, leave managers can adapt their workflow and data management systems to expand the information they collect. This will allow the formal capture of important case information within a leave database. As the case data grows, you will develop the capability to predict high-risk cases and extract other analytic insights.

Another path, as part of your expanded leave management data system, is the use of natural language processing (NLP) to extract diagnosis, procedure, and pharmaceutical information from the free text notes of a claim. At MDGuidelines, we have begun to apply Unified Medical Language System (UMLS) lexicons such as SNOMED and RxNorm to systematically code and group free text.4 For example, if a leave manager writes in the notes of a claimant, “Case has LBP, currently on hydrocodone,” then an NLP system using SNOMED would be able to extract from that sentence that the patient has low back pain and is taking an opioid medication. Further, these UMLS groupings make it much easier to integrate leave management systems with electronic health records, which will increase our capabilities to link to medical information and improve our predictive analytics platforms.


This small data project revealed that we need to align nurse case managers’ knowledge about each case with what an analyst can learn from the data. With this integration, the leave management industry will actually be able to extract insights from the data. Although larger organizations are leading this development, employers of all sizes can now use tools from vendors and the UMLS to extract valuable insights that can improve case management practices and outcomes.


  1. Readers can view and download ReedGroup’s “Case Review Extraction Form” at
  2. Manek NJ, AJ MacGregor. Epidemiology of Back Disorders: Prevalence, Risk Factors, and Prognosis. Current Opinion in Internal Medicine. 4(3):324-330. 2005. doi:10.1097/01.bor.0000154215.08986.06.
  3. Nichols GA, K Reinier, SS Chugh. Independent Contribution of Diabetes to Increased Prevalence and Incidence of Atrial Fibrillation. Diabetes Care.32(10):1851-1856. 2009. doi:10.2337/dc09-0939.
  4. U.S. National Library of Medicine. Unified Medical Language System. 2016. Retrieved at