Open Wound Research

How many mobile wound care providers are there?

Zwelithini Tunyiswa — Tue, 22 Jul 2025 00:00:00 GMT

I was speaking to my co-founder of Open Wound Research, Ryan Dirks, and an interesting question came up between us. Just how many mobile wound care providers are there? By providers, I mean individual providers versus mobile wound care groups. This got my curiosity going, and I thought I would work in public to estimate this number. I am going to use CMS Medicare Utilization by Provider and Service file. Unfortunately, the data is only available up to calender year 2023.

Let’s load the file, and get started.

Code

The Braden Scale as a Statistical Model

Zwelithini Tunyiswa — Thu, 12 Sep 2024 00:00:00 GMT

I just did a great podcast with Rob Fraser from SWIFT, and we began talking about Pressure Ulcer risk, the Braden Scale, and machine learning. It got me thinking about how to talk about the Braden Scale as modeler. Below is the result from cobbling together some notes that I wrote to myself over the past couple of months.

Raw Braden Scores are crude homogenizations

In modeling, we simplify reality to make sense of complex phenomena. The Braden Scale, like many clinical assessment tools, is one such simplification. It uses a set of subscales to capture the different dimensions of a patient’s risk of developing pressure ulcers. This is an attempt to formalize a process that is inherently complex and probabilistic.

A score of 10 might be useful as a crude signal—lower scores indicate higher risk—but what we really care about is how the patient got that score. The total score hides the deeper, multi-dimensional structure of the data. To better understand scores, we need to decompose them.

Thinking Hierarchically about the Braden Scale

My preference is to think in terms of the heuristic of a hierarchical model. To wit, a Braden score is composed of six subscales—sensory perception, moisture, activity, mobility, nutrition, and friction/shear. Each subscale captures a different source of risk, but the real danger often comes from the interactions between these subscales, not just their individual effects.

For instance, if a patient has a low score in “sensory perception” (they can’t feel pressure), and also a low score in “mobility” , this combination is much more dangerous than either subscale by itself. These two factors interact in a multiplicative, not additive, way. The inability to feel discomfort plus the inability to move creates a joint risk of skin damage that grows exponentially, like a latent process waiting to erupt.

If we model this in a Bayesian framework, we might expect an interaction term in our model —where the product of sensory perception and mobility contributes disproportionately to the likelihood of pressure ulcers. In the current additive framework of the Braden Scale, this interaction is hidden. So, while two patients might have the same total score, the underlying structure of risk could be very different.

Aggregation obscures important information

Aggregation —like summing different dimensions into a single score—can mask important variation and nonlinear relationships. A total score of 10 assumes that the risk from each subscale is independent and contributes equally to the total. But this is rarely true in real-world systems. Each subscale might have nonlinear impacts that are contingent on the levels of the others. For instance, the effect of “friction and shear” on skin damage is probably much worse if the patient’s “mobility” score is low, because they can’t reposition to relieve the pressure. But this interaction isn’t captured by the simple sum.

Through the frosted looking-glass

In a Bayesian hierarchical model, we would ideally model these subscales not as independent sources of risk, but as “partial observables” of a latent process—perhaps something like “skin breakdown vulnerability”, or “skin frailty”. This latent variable would then be influenced by a combination of subscales, each with its own weight and interaction effects. Rather than reducing everything to a single number, we would retain the structure that allows for these nuanced relationships to emerge.

The Braden Scale might be useful for (many disagree) predicting the overall risk of pressure ulcers, but it doesn’t necessarily explain the mechanisms underlying that risk. This is a key distinction. The explanatory power is what helps clinicians understand how and why risk arises in a given patient. A simple score can tell you that two patients are at risk, but it doesn’t explain why they are at risk in different ways.

For instance, if Patient 1 is at risk due to impaired “sensory perception” and “mobility”, they might require frequent repositioning and sensory interventions. If Patient 2 is at risk due to “moisture” and “nutrition”, interventions would focus more on skin care and dietary improvements. These are very different clinical responses, but they both could be hidden by a total score of 10, which offers no insight into the causal mechanisms at play.

Down the Bayesian Rabbit-Hole

In the Bayesian framework, after fitting a model, we use the posterior distribution to generate new data and see how well it matches the real-world data. In the case of the Braden Scale, if we were modeling pressure ulcer risk, we’d want to check whether patients with the same total score (like two patients with a score of 10) have similar outcomes in terms of pressure ulcer development. If one patient develops ulcers much faster than another despite the same total score, this suggests the model (or scale) is missing important aspects of the data structure. In this case, it is probably the interactions between subscales.

We could place priors on the subscales based on expert knowledge or empirical data about which factors are most important for pressure ulcer development. Instead of assuming equal contributions from each subscale, we could allow the model to learn which combinations of factors are most dangerous. Over time, this would produce a better, more nuanced model—one that doesn’t rely on the misleading simplicity of a total score.

Saying what I said – for the lazy readers

The Braden Scale is a useful tool, but like all models, it is an abstraction. When we collapse six subscales into one number, we lose critical information about how these subscales interact and influence the risk of pressure ulcers. What we need is a model that respects the hierarchical structure of the data and allows for interaction effects between subscales. By focusing on these interactions and latent processes, we can move beyond simple predictions and develop a deeper understanding of the mechanisms of risk. This is the path to better, more effective clinical interventions, rather than relying on a single, oversimplified score.

Simulating Wound Data is Easier Than You Think

Zwelithini Tunyiswa — Mon, 06 May 2024 00:00:00 GMT

As a modeler, I oftentimes simulate wound care data to test models prior to fitting those same models to real data. Why would one do this? With simulated data one know exactly what the ground truth or parameters are, and can test how close the model gets to estimating the true values. With real data, the ground truth can never truly be known with absolute certainty. It can only be estimated. Hence, simulated data is a playground for models, before they make their way to the complexity of Real World Evidence.

I mentioned simulation of data to an aquaintance the other day, and they said it must be pretty difficult. I assurred him it was very straightforward, and in fact one could generate very realistic longitudinal wound care data do so with a couple of lines of code and some very simple distributional assumptions.

First let’s load the Python libraries we will need.

import altair as alt # our visualization library
import polars as pl # our dataframe library
#pl.Config.set_tbl_rows(24) # allow the dataframe to print 24 rows without clipping
import pymc as pm # the library that we will use to simulate data

Next let’s define the number of wounds we want, the starting area of the wounds (drawn from a Gamma distributution), and the innovations of the wounds, or how the wounds change week to week (drawn from a StudentT distribution). We will generate 12 weeks of data for 20 wounds.

number_of_weeks = 11 # this is actually 12 weeks as we have our starting area + 11 more weeks
number_of_wounds = 20 
random_seed = 1234 # this allows the simulations to to be replicated, otherwise each draw will be totally random


starting_areas = pm.Gamma.dist(mu=7.5, # the average area
                               sigma=3.5 # the dispersion paramater
                               )

wound_trajectories_definition = pm.RandomWalk.dist(init_dist=starting_areas,
                                        innovation_dist=pm.StudentT.dist(nu=3, mu=-.3, sigma=1), # the innovations with a skew toward healing
                                        steps=number_of_weeks
                                        )

wound_trajectories_draws = pm.draw(vars=wound_trajectories_definition, 
                                   draws=number_of_wounds, 
                                   random_seed=random_seed
                                   )

Let’s take a glance is returned by the simulation. We return two arrays of data, which represent the trajectories over 12 weeks of two of the 20 wounds.

print (wound_trajectories_draws[:][:2])

[[ 8.31573535 10.08365913  8.60467865  7.61161127  7.803105    4.94285792
   5.55018078  4.13138653  4.03536057  2.57860449  2.83068657  0.77546271]
 [ 7.35366688  8.09203598  6.65627211  7.15055862  7.69705732  8.88698897
   7.66514015  7.5456083   6.85247688  3.65505568  2.75086306  2.54986529]]

Looking at array data is not terribly pleasant, so let’s convert this to a tabular format.

df = (pl
      .DataFrame(wound_trajectories_draws, orient='row') # load the array data into a dataframe
      .with_row_index(name='wound_id') # generate the wound_id
      .with_columns(wound_id = pl.concat_str([pl.lit('wound_'), pl.col.wound_id.cast(pl.String).str.zfill(3)])) # clean up the wound_id column
      .melt(id_vars=['wound_id'], variable_name='time', value_name='area') # make the data long
      .with_columns(pl.col.time.str.split('_').list.last().cast(pl.Int16).add(1)) # clean up the time column
      .with_columns(pl.when(pl.col.area.le(0)).then(0).otherwise(pl.col.area).alias('area')) # clean up area column for when wound heals
      .sort(by=['wound_id', 'time']) # some sorting to make sense of things
      )

df.to_pandas().head(24) # show the same 2 wounds as in the array above

	wound_id	time	area
0	wound_000	1	8.315735
1	wound_000	2	10.083659
2	wound_000	3	8.604679
3	wound_000	4	7.611611
4	wound_000	5	7.803105
5	wound_000	6	4.942858
6	wound_000	7	5.550181
7	wound_000	8	4.131387
8	wound_000	9	4.035361
9	wound_000	10	2.578604
10	wound_000	11	2.830687
11	wound_000	12	0.775463
12	wound_001	1	7.353667
13	wound_001	2	8.092036
14	wound_001	3	6.656272
15	wound_001	4	7.150559
16	wound_001	5	7.697057
17	wound_001	6	8.886989
18	wound_001	7	7.665140
19	wound_001	8	7.545608
20	wound_001	9	6.852477
21	wound_001	10	3.655056
22	wound_001	11	2.750863
23	wound_001	12	2.549865

Great. This is makes the wound data easier to read. Let’s go one further and visualize the trajectories for for all 20 wounds.

The simulation of authentic looking data was accomplished in less than 15 lines of code. Further complexity could be added to insert censoring (interval, left, or right), truncation, lower and upper bounds of starting area, etc.

Simulation is an absolute must in the toolkit of any wound-care researcher.

An Initial Rough Estimate of the Potential Cost to the CAMPS Industry of the new CMS LCDs

Zwelithini Tunyiswa — Tue, 30 Apr 2024 00:00:00 GMT

All eight CMS Medicare Administrative Contractors (MACs) promulgated proposed concurrent Local Coverage Determinations (LCD’s) on April 25th for Skin Substitutes. In all but 2 MACs, only 15 products would be covered, and coverage would be limited to 4 applications in a 12-week epoch. The Alliance of Wound Care Stakeholders has written a great overview of the nature and scope of the LCD’s.

An initial thought that I had upon review of the LCD’s this past weekend is the potential cost to the industry. What could the potential impact be? So, I decided to do a a rough estimate using some internal OWR data and simple assumptions.

Proceed with caution!

This is a thought experiment and not a study, statement of fact, or an opinion about the LCDs.

At OWR we have a CAMPS Medicare claim dataset that we put together for research and application development. Here, we used the data for the calendar year 2022. We found the 15 products, and their HCPCS codes. The products/codes that are covered are as follow:

	hcpcs_cd	hcpcs_cd_short_description	lcd_status
0	Q4159	AFFINITY1 SQUARE CM	covered
1	Q4168	AMNIOBAND, 1 MG	covered
2	Q4151	AMNIOBAND, GUARDIAN 1 SQ CM	covered
3	Q4101	APLIGRAF	covered
4	Q4122	DERMACELL, AWM, POROUS SQ CM	covered
5	Q4106	DERMAGRAFT	covered
6	Q4187	EPICORD 1 SQ CM	covered
7	Q4186	EPIFIX 1 SQ CM	covered
8	Q4145	EPIFIX, INJ, 1MG	covered
9	Q4128	FLEXHD/ALLOPATCHHD/SQ CM	covered
10	Q4133	GRAFIX STRAVIX PRIME PL SQCM	covered
11	Q4107	GRAFTJACKET	covered
12	Q4113	GRAFTJACKET XPRESS	covered
13	Q4105	INTEGRA DRT OR OMNIGRAFT	covered
14	Q4103	OASIS BURN MATRIX	covered
15	Q4124	OASIS TRI-LAYER WOUND MATRIX	covered
16	Q4102	OASIS WOUND MATRIX	covered
17	Q4110	PRIMATRIX	covered
18	Q4121	THERASKIN	covered

We then set the LCD coverage status of products/HCPCS codes in the dataset. A sample of the final dataset looks like this:

	hcpcs_cd	hcpcs_cd_short_description	lcd_status	units	payments	payment_per_unit
0	Q4188	AMNIOARMOR 1 SQ CM	non-covered	52,494.0	24,675,938.39	470.07
1	Q4169	ARTACENT WOUND, PER SQ CM	non-covered	148,464.5	39,159,374.52	263.76
2	Q4210	AXOLOTL GRAF DUALGRAF SQ CM	non-covered	47,228.6	29,083,601.85	615.8
3	Q4105	INTEGRA DRT OR OMNIGRAFT	covered	168.0	24,435.53	145.45
4	Q4205	MEMBRANE GRAFT OR WRAP SQ CM	non-covered	41,491.0	10,068,052.33	242.66

We can then roll up the numbers by covered vs. non-covered, in terms of units and Medicare payments.

	lcd_status	payments
0	covered	121,836,612.57
1	non-covered	1,134,254,167.08

We now need an estimate of the ratio of diabetic and venous ulcers claims/payments for Medicare beneficiaries that we can apply to the non-covered cohort (remember the LCD’s are limited to diabetic and venous ulcers). Fortunately, Marissa Carter et al wrote a wonderful paper in 2023¹ that is fount of information regarding Medicare wound-type reimbursement. We can take a look at Figure 4 in that paper.

If we sum the diabetic and venous wound-types (the principal diagnosis column) we get that they represent 0.2459 of the total. If we apply that proportion to the non-covered cohort - the math looks like this:

$278,913,099.68² … is what these LCD’s could potentially cost the CAMPS industry in Medicare payments.

That is quite a number…

Footnotes

Carter MJ, DaVanzo J, Haught R, Nusgart M, Cartwright D, Fife CE. Chronic wound prevalence and the associated cost of treatment in Medicare beneficiaries: changes between 2014 and 2019. J Med Econ. 2023 Jan-Dec;26(1):894-901. doi: 10.1080/13696998.2023.2232256. PMID: 37415496.↩︎
This number is probably conservative as we have not factored in that coverage would be limited to 4 applications in a 12-week epoch in covered products↩︎

An Overview of Nonequivelent Group Study Designs in Wound Care

Zwelithini Tunyiswa — Sun, 28 Apr 2024 00:00:00 GMT

A Quick Intro to Nonequivelent Group Designs

Much of wound-care research is not randomized control studies, but rather quasi-experimental studies. The most commonly used quasi-experimental design is the nonequivalent group design (NEGD)¹.

According to Reichardt, “in the nonequivalent group design, groups of participants are non-randomly assigned to receive either a treatment or a comparison condition … the nonequivalent group design is akin to a between-groups randomized experiment in that both designs draw comparisons between participants who receive different treatment conditions”².

Notation for Study Design Diagrams

The notation for the Study Design Diagrams is straightforward:

NR = Not Randomized

X = Treatment | X_∀ = A specific treatment

O = Observation | O_n = Observation Number

Control Groups in Wound Care

I refer to control groups below in the description of NEGD’s. However, pure controls don’t really (at least purposefully!) exist in wound care. If a patient has a wound they are generally being treated using Standard-of-Care (SOC). So really SOC is the control. However, this takes us down a rabbit hole as there manifold SOC’s in the real world. Although there are national guidelines and the like, a SOC in a wound care center may differ significantly from a SOC in a post-acute mobile practice, or a home-care practice. Even SOC may differ between providers in the same practice. However, regardless of the noise in definition, control can be read to mean SOC.

In wound-care, these, in my opinion, are the most prevalent NEGD’s:

Posttest-only Nonequivelent Group Design

This is the simplest NEGD. This design has no pretest (or observation before treatment for neither the control nor the control group). The treatment is administered the treatment group and the only measurements collected are at the end of the study. The control and treatment groups are only compared on post endpoint differences. This design lacks a pretest, so we are unable to assess the magnitude of change or potential bias at pretest.

Posttest-only Nonequivelent Group Design
NR		X	O₁
NR			O₁

The Pretest-Posttest Nonequivelent Group Design

This is the most common NEGD. A Pre-test measurement is taken in both groups, a treatment administered in the treatment group, and endpoints assessed at the end of the study. This setup allows allows one to asses the pre and post group differences, and allows for the assessment of change intra and inter group.

The Pretest-Posttest Nonequivelent Group Design
NR	O₁	X	O₂
NR	O₁		O₂

Another variation of this design is to add pre observations, so that the diagram looks something like this:

The Pretest-Posttest Nonequivelent Group Design (with additional pre-test observation(s))
NR	O₁	O₂	X	O₃
NR	O₁	O₂		O₃

This extra observation allows for a quasi run-in period that allows the researcher to assess for change between pre-test observations (was one group already healing/worsening before O₂?). One could also add post-test observations to assess relapse/recurrence. That could look something like this:

The Pretest-Posttest Nonequivelent Group Design (with additional post-test observation(s))
NR	O₁	X	O₂	O₃
NR	O₁		O₂	O₃

The Pretest-Posttest Alternative Treatments Nonequivelent Group Design

A variation on the the Pretest-Postest NEGD to switch out the control for an alternative treatment. Hence there are two treatment groups. The same addition of pre-test and post-test observations as in Pretest-Postest NEGD.

The Pretest-Posttest Alternative Treatments Nonequivelent Group Design
NR	O₁	X_A	O₂
NR	O₁	X_B	O₂

Multiple Treatments and Controls with Pretest

Another variation of the Pretest-Postest NEGD is to simply add another treatment group, which would look like this. The same addition of pre-test and post-test observations as in Pretest-Postest NEGD.

Multiple Treatments and Controls with Pretest
NR	O₁	X_A	O₂
NR	O₁	X_B	O₂
NR	O₁		O₂

Notwithstanding concerns around bias and internal validity, NEGD are necessary in wound care and form the underpinning of the notion of Real World Evidence (RWE). Understanding their structure, and the appropriate tools for their analysis (more on that to come) is important as we advance wound-care research.

Footnotes

In wound care, it is more commonly called an observational study design, the nomenclature made popular by Cochran and Rosenbaum. However, my preference is for the nomenclature from Campbell, which is more contemporary, and I believe captures the idea of what is actually occurring more accurately (we are, more often than not, not really just observing, are we?).↩︎
Reichardt, C. S., & Little, T. D. (2019). Quasi-experimentation: A guide to design and analysis. The Guilford Press. , p. 140↩︎