Introduction

The cost of living varies by state as well as the financial well-being.  It follows that the CMS reimbursement rates will differ regionally as well. These regional differences will need to be taken in account to improve the accuracy per state.

This is complicated by the lack of price transparency.  Robert Zirkelbach, from the America’s Health Insurance Plans association states “There’s very little transparency out there about what doctors and hospitals are charging for services. Much of the public policy focus has been on health insurance premiums and has largely ignored what hospitals and doctors are charging.” (Meier 2013)

R Libraries

A note about the R libraries that proved useful for this blog post. The sqldf package allows one to use SQL for composing new data frames. Josh Mills (2012) wrote a good blog post of its use.

avg.cost.by.state=sqldf("select [Provider.State], 
sum([Average.Estimated.Submitted.Charges]*[Outpatient.Services])
/sum([Outpatient.Services]) as stateCharges,
sum([Average.Estimated.Submitted.Charges]/[Average.Total.Payments]*[Outpatient.Services])
/sum([Outpatient.Services]) as overage,
sum([Average.Total.Payments]*[Outpatient.Services])
/sum([Outpatient.Services]) as [statePayments]
from statedf group by [Provider.State]")

Steve Bronder (2014) provides an example of the statebins package.  The appealing nature of statebins is that each state is given the same physical space on the screen regardless of its actual landmass. This allows for comparing states to be easier and more intuitive.

statebins_continuous(avg.cost.by.state,
 "Provider.State", "statePayments",
 legend_title="legend", font_size=3, 
 brewer_pal="PuRd", text_color="black", 
 plot_title="Weighted Avg State Payments"
 , legend_position="bottom", 
 title_position="top") + guides(fill = guide_colorbar(barwidth = 10, barheight = 1))

Reimbursement Rates by State

The data used is from Centers for Medicare & Medicaid Services (2015) “Outpatient Charge Data CY 2013” which provides charge and reimbursement amounts for states and a selected set of hospitals.  However, unfortunately no data is provided for Maryland.

The following chart shows the weighted average state payments (CMS reimbursements).  The average is weighted by the number of services performed of the various service estimated hospital-specific charges for 30 Ambulatory Payment Classification (APC) Groups paid under the Medicare Outpatient Prospective Payment System (OPPS) for Calendar Year (CY) 2013. Notice that Wyoming has the largest per service reimbursements.

A charge master is a list detailing the official rate charged by a hospital for individual procedures, services, and goods. “The charge master is used to generate each hospital invoice”. (Google)

Each hospital uses their own charge master to bill CMS, however:

Medicare does not actually pay the amount a hospital charges but instead uses a system of standardized payments to reimburse hospitals for treating specific conditions. Private insurers do not pay the full charge either, but negotiate payments with hospitals for specific treatments. Since many patients are covered by Medicare or have private insurance, they are not directly affected by what hospitals charge. (Meier 2013)

Notice that Alabama and New Jersey show larger differences between what they billed and what they were reimbursed.

Drilling down to the hospital shows more clearly the metropolitan areas.

Meier reports that experts say it is likely that the people with little or no insurance are getting hit with extremely high hospitals bills that may bear little connection to the cost of treatment.  “If you’re uninsured, they’re going to ask you to pay,” said Gerard Anderson, the director of the Johns Hopkins Center for Hospital Finance and Management. (Meier 2013)

Are the bills higher in states which have low insurance rates?  Is it a form of charging excessive charges for those who are “self pay” patients?

Comparing the two datasets side by side, one can see some similarities.

The correlation between average charges and the uninsured% is 36%.

Performing regression analysis it appears that perhaps up to 11% of the variance in pricing may be predictable by the uninsured population.

The coefficient of determination (denoted by R²) is a key output of regression analysis. It is interpreted as the proportion of the variance in the dependent variable that is predictable from the independent variable. (Stat Trek)

Conclusion

A potential patient could save a large of amount of money by shopping around for medical services if the pricing was more transparent. Unfortunately, those who are uninsured are affected the most. Using this data, it may prove advantageous to travel to those states with a low overage compared to the CMS reimbursement. (For example, Minnesota.)

References

Centers for Medicare & Medicaid Services (2015) “Outpatient Charge Data CY 2013” https://www.cms.gov/Research-Statistics-Data-and-Systems/Statistics-Trends-and-Reports/Medicare-Provider-Charge-Data/Outpatient2013.html

Introduction

The challenge being examined here in healthcare is to make hospitals pricing transparent.  Hospitals do not have a list of high level services with prices that most people would understand.  High level services are billed out as claims of individual charges.  Those charges are encoded as Current Procedural Terminology (CPT) codes. If there is a manner in which groups of codes are commonly performed in encounters, those separate groups could act as the list of services.

If one considers a claim a document and a code a word, one can use bag-of-words and the vector-space-model techniques as methods for analyzing claims.

Kunwar (2013) divided clustering techniques into the four broad categories:

1. Flat clustering (Creates a set of clusters without any explicit structure that would relate clusters to each other; It’s also called exclusive clustering)
2. Hierarchical clustering (Creates a hierarchy of clusters)
3. Hard clustering (Assigns each document/object as a member of exactly one cluster)
4. Soft clustering (Distribute the document/object over all clusters)

In the previous blog post, LDA topic classification was explored.  LDA analysis is considered flat and soft clustering.  The soft clustering produced topics that could allow claims to be in more than one topic at a time.  At first this may seem ideal since a hospital encounter could involve removing the gall bladder and the appendix.  However, the results produced washed-out clusters and in this author’s opinion were only suitable for high-level categories.

In this post, an exploration of K-means analysis is performed.  K-means is a method of flat and hard clustering. In addition, two methods of vector creation will be explored term frequency (TF) and Term Frequency / Inverse Document Frequency (TFIDF).

For this week’s blog post, Apache Mahout and Anaconda Python Jupyter are utilized.

How Many Clusters?

Using the Elbow criteria (looking for an elbow in a graph), if one looks at the average weight of the word (code) when using TF vectors, the number of clusters should be above 30 due to the average term weight (green line) crossing 1.0.

When using TFIDF with normalization and with prior knowledge that 90% of the claims have less than 8 claim lines, the top red line crosses 90% around 200.  So the number of topics should be greater than 200.

Can, F.; Ozkarahan (1990) theorize that the number of clusters in text databases can be estimated by

number of clusters estimate=(m*n)/t

where:

• m: number of claims
  • Via: select count(clm_id) from claims
• n: number of unique codes
  • Via: select count(distinct hcpcs_cd) from claimscpt
• t: count of codes present in every claim
  • Via: select count(*) from claimscpt

m=744995
n=5618
t=2831888
(m*n)/t
1477.94

Thus using the above methods, the number of clusters should be around 1477.

Interestingly in a previous blog post, I theorized that an indicator of a “procedure” CPT code as one that has non-zero work units as defined by CMS in its relative value units calculation.

American College of Radiation Oncology states: Physician work RVU – The relative level of time, skill, training and intensity to provide a given service. A code with a higher RVU work takes more time, more intensity or some combination of these two. Some radiation oncology codes, such as treatment codes, have no associated physician work.

Using that criteria, the number of services came out to be 1603.

Cluster Interpretation

Mahout clustering goes through a series of steps to output the clusters.  As part of those steps one must choose TF or TFIDF for the vector type.

mahout seq2sparse -i out/sequenced \
    -o out/sparse-kmeans -wt TF --maxDFPercent 100 --namedVector

…

mahout kmeans \
    -i out/sparse-kmeans/tf-vectors/ \
    -c out/kmeans-clusters \
    -o out/kmeans \
    -dm org.apache.mahout.common.distance.CosineDistanceMeasure \
    -x 10 -k $i -ow --clustering 

Term Frequencies

The output from Mahout’s clusterdump produces a list of terms belonging to a topic. What is the exact meaning of the numerical value below from clusterdump?

Top Terms: 
 g0206 => 0.22002189800292146

In the above 0.22 is the center of the centroid for that term. If the term vectors were created with term frequencies, could one interpret this as the average occurrence of the “G0206” in the documents that are considered part of the cluster. In this case, “G0206” could appear in 22% of the docs in that cluster or more likely that the average count of G0206 in those claims is 0.22.

Looking further, one can see codes like so:

Top Terms:
71010   =>  3.5612545939557902

So it is best to interpret this as the average count of “71010” in the claims that are in that set of documents.

Given the radius values from the cluster dump output, one could consider that the range of unit occurrences of the individual codes.

Using Term Frequencies has a desired effect of being the unit count and thus lends itself to calculating cluster reimbursement costs.

TF-IDF

When one uses tf-idf weighting, it may be best to normalize the output weights by creating a proportion of evidence via Wi=Wi/sum(W). The tf-idf weight by itself is not directly interpretable except as sense of level of evidence.  This appears to be common practice as for example, the normalization of the output is done by the GenSim and LDA Python libraries.

Cluster Results

A note about pricing…

Remember with term frequency, the weight associated with each term can be interpreted as the unit quantity.  That is because the code word was repeated for each unit.

With term frequency/inverse document frequency, the weight associated is more of a heuristic of evidence of the prevalence of the code word.

In the TF cluster results,  the weight can range from 0 to N (unit quantity from term frequency).

price = (cptBasePrice + cptUnitPrice * max(1,weight) ) * min(1,weight)

Thus with the above, at least 1 unit will be used and if the weight is less than 1, it will be treated as a probability of the code being used.

An  alternative method is for TFIDF normalized weights probabilities to be used solely for the prevalence of the code while using the average reimbursement price as the normalized weight can range from 0 to 100% (evidence of its presence from TFIDF).

The best method to test the methods would be to measure the error (RMSE) from the actual and predicted costs of each claim.

TF-IDF K=1477

The following clusters (K=1477, eps=0.10, with normalization) look sound as for the code grouping themselves, but the expected value of Medicare reimbursement seems rather low.

Notice that the following three clusters are for UTIs, mammograms and therapies.

Topic  2
87086 0.35 $ 6.49 Urine culture/colony count
81001 0.31 $ 2.93 Urinalysis auto w/scope
     Total $9.42

Topic  12
76645 0.60 $38.77 breast ultrasound
77051 0.13 $ 1.56 Computer dx mammogram add-on
     Total $40.33

Topic  41
97110 0.20 $ 5.08 Therapeutic exercises
97116 0.17 $ 3.46 Gait training therapy
97530 0.17 $ 4.30 Therapeutic activities
97140 0.15 $ 3.30 Manual therapy 1/> regions
     Total $16.14

 TF K=1477

The following clusters (K=1477, eps=0.10, without normalization) look sound as for the code grouping themselves as well and notice the reimbursement rates are higher.

Topic 954 matches topic 12 above.

Topic  954
76645 1.30 $64.38 breast ultrasound
77051 0.37 $ 4.43 Computer dx mammogram add-on
     Total $68.81

Looking at the clusters, occasionally other groupings are found with the same codes and in this scenario, perhaps a cancerous tumor was detected.

Topic  101
77051 1.18 $11.59 Computer dx mammogram add-on
88305 1.03 $36.38 Tissue exam by pathologist
76645 0.10 $ 6.76 breast ultrasound
     Total $54.73

K Choice Sensitivity

TF IDF

Inspecting the sensitivity of choosing K, below is K=4 TFIDF with normalization. With eps=0.1, sometimes no codes would be displayed in those cases the “top line” item is displayed.

eps 0.1  normalize=True

Topic  0
97110 0.04 $ 1.92 Therapeutic exercises
     Total $1.92

Topic  1
36415 0.12 $ 0.81 Routine venipuncture
     Total $0.81

Topic  2
80053 0.05 $ 1.94 Comprehen metabolic panel
     Total $1.94

Topic  3
99211 0.21 $11.17 Office/outpatient visit est
     Total $11.17

For the sake of completeness, below is K=4 without normalization revealing the TF-IDF weights (centroid centers). The results are less than satisfying and reveal a sensitivity to choosing the epsilon cut-off.  (The TF-IDF heuristic has no meaningful intrinsic value to 0.1 verses 0.2 per se.) Only the first topic shown.

Topic  0
97110 0.74 $36.97 Therapeutic exercises
q4081 0.58 $ 0.00 Epoetin alfa, 100 units esrd
a4657 0.57 $ 0.00 Syringe w/wo needle
90999 0.57 $69.88 Dialysis procedure
j2501 0.44 $ 0.00 Paricalcitol
97530 0.38 $19.49 Therapeutic activities
99213 0.37 $23.86 Office/outpatient visit est
97112 0.32 $13.58 Neuromuscular reeducation
97116 0.32 $10.50 Gait training therapy
97140 0.30 $11.98 Manual therapy 1/> regions
99212 0.28 $19.62 Office/outpatient visit est
j1756 0.22 $ 0.00 Iron sucrose injection
j1270 0.22 $ 0.00 Injection, doxercalciferol
77052 0.21 $ 2.23 Comp screen mammogram add-on
99214 0.20 $14.37 Office/outpatient visit est
71020 0.20 $ 7.98 Chest x-ray 2vw frontal&latl
g0202 0.20 $ 0.00 Screeningmammographydigital
g0283 0.19 $ 0.00 Elec stim other than wound
97001 0.17 $ 9.79 Pt evaluation
97535 0.16 $ 8.40 Self care mngment training
90945 0.15 $ 9.66 Dialysis one evaluation
99283 0.15 $12.35 Emergency dept visit
97035 0.15 $ 1.97 Ultrasound therapy
87086 0.13 $ 2.41 Urine culture/colony count
81001 0.12 $ 1.15 Urinalysis auto w/scope
85610 0.12 $ 1.89 Prothrombin time
88305 0.12 $ 7.46 Tissue exam by pathologist
92526 0.12 $ 9.19 Oral function therapy
g0008 0.11 $ 0.00 Admin influenza virus vac
90658 0.10 $ 1.56 Flu vaccine 3 yrs & > im
85025 0.10 $ 2.53 Complete cbc w/auto diff wbc
93798 0.10 $ 3.87 Cardiac rehab/monitor
     Total $302.69

 TF

With term frequency, the value of the weight has a meaning. It is the mean average units for that CPT code.

Looking at the results of a small set of clusters (K=4, eps=0.1  TF without normalization), the results are still sensible but just not exhaustive.  This is good news as it shows this method to be not be sensitive to the size of K to produce coherent and sensible clusters.

eps 0.1

Topic  0
q4081 0.39 $ 0.00 Epoetin alfa, 100 units esrd
a4657 0.39 $ 0.00 Syringe w/wo needle
90999 0.37 $45.51 Dialysis procedure
j2501 0.23 $ 0.00 Paricalcitol
80053 0.18 $ 6.73 Comprehen metabolic panel
85610 0.13 $ 1.95 Prothrombin time
99213 0.10 $ 6.62 Office/outpatient visit est
     Total $60.82

Topic  1
97110 3.00 $65.63 Therapeutic exercises
97530 1.01 $25.22 Therapeutic activities
97116 0.78 $15.53 Gait training therapy
97112 0.75 $17.83 Neuromuscular reeducation
97140 0.67 $14.51 Manual therapy 1/> regions
g0283 0.39 $ 0.00 Elec stim other than wound
97535 0.27 $ 6.56 Self care mngment training
97035 0.25 $ 2.98 Ultrasound therapy
97001 0.22 $13.08 Pt evaluation
97150 0.11 $ 1.60 Group therapeutic procedures
     Total $162.93

Topic  2
87086 0.15 $ 2.78 Urine culture/colony count
81001 0.13 $ 1.23 Urinalysis auto w/scope
87186 0.11 $ 1.82 Microbe susceptible mic
99283 0.10 $ 8.49 Emergency dept visit
     Total $14.32

Topic  3
36415 0.78 $ 4.70 Routine venipuncture
85025 0.37 $ 9.12 Complete cbc w/auto diff wbc
80061 0.18 $ 4.81 Lipid panel
85610 0.18 $ 2.74 Prothrombin time
80053 0.14 $ 5.14 Comprehen metabolic panel
93005 0.13 $ 3.81 Electrocardiogram tracing
80048 0.13 $ 3.36 Metabolic panel total ca
84443 0.11 $ 3.86 Assay thyroid stim hormone
     Total $37.54

Conclusion

K-Means looks to be a sound approach with TF vectors producing better results than TF IDF in this study for this use case for several reasons:

1. Term Frequencies act as a stand in for the number of units since each code was repeated for each unit.  This makes the weights interpretable.
2. Each code has a level of meaning.  No code acts as a “stop word”.
3. The choice of K with TF vector clusters is less sensitive to thresholds and cut offs for topic word selection.

Next read:

State of Care

References

American College of Radiation Oncology (2009) Introduction to Relative Value Units and How Medicare Reimbursement in Calculated http://www.acro.org/washington/rvu.pdf

Apache Mahout (n.d.) “Cluster Dumper – Introduction” https://mahout.apache.org/users/clustering/cluster-dumper.html

Chris Busch (2016) “Mahout clusterdump top terms meaning” http://datascience.stackexchange.com/questions/10987/mahout-clusterdump-top-terms-meaning

Samir Kunwar (2013) “Text Documents Clustering using K-Means Algorithm” http://www.codeproject.com/Articles/439890/Text-Documents-Clustering-using-K-Means-Algorithm

Can, F.; Ozkarahan, E. A. (1990). “Concepts and effectiveness of the cover-coefficient-based clustering methodology for text databases”. ACM Transactions on Database Systems 15 (4): 483. doi:10.1145/99935.99938  https://en.wikipedia.org/wiki/Determining_the_number_of_clusters_in_a_data_set#Finding_Number_of_Clusters_in_Text_Databases

Revision: Sat April 2 1:45 AM

Introduction

Every time someone visits the hospital a claim is generated. On each of those claims is a set of codes that signifies the procedures that were performed. Let us think of every single procedure as a word in a document and the claim itself is a document. Now if we imagine a set claims as separate documents we can explore topic analysis to determine which procedures should go with other procedures. This could be useful for determining proper coding of medical claims or for doing research.

LDA

Latent Dirichlet Allocation (LDA) gives an output for each word in the dictionary.  No word in the dictionary has a zero result. So in some sense all words are in a topic.

So a question is what words to keep in each topic?

There are several methods that come to mind:

• Choose the top N words.
• Choose the top words above a cut-off threshold (epsilon)
• Choose the top words whose lambda values sum up over a threshold

Let us study the Reuters data set to explore these cut off methods.

Top N Words

Looking at the used the code in the blog post Topic modeling with latent Dirichlet allocation in Python, in that code, the author shows the top 8 words in each topic, but is that the best choice?

Notice above that even with 200 topics, the top 8 words only account for about 35% of the topic on average.  Unless there is a compelling need for a short topic phrase, 8 seems arbitrary.

Epsilon Cut-Off

For each topic distribution, each word has a probability and all the words probabilities add up to 1.0 when the word lambda values are normalized.

I wrote this code to print out down to an epsilon threshold:

eps=0.01
for i, topic_dist in enumerate(topic_word):         
    wordindex=np.argsort(topic_dist)[::-1] #rev sort
    w=topic_dist[wordindex] ## this is the length of all the unique words 4258
    words=[np.array(vocab)[wordindex[j]] for j in range(min(n_top_words,len(wordindex))) if w[j]>eps ]
    weights=['{:.3f}'.format(w[j]) for j in range(min(n_top_words,len(wordindex))) if w[j]>eps ]
    print('Topic {}: {}; {}'.format(i, ', '.join(words),', '.join(weights)))

What is the best eps to choose? Is it best to use the max probability value in each topic and base a cut-off from that?

In the datasets, sometimes an eps of 0.01, actually creates a word-less topic!!

Playing with different numbers of topics, I noticed that if I have 2 topics with the load_reuters data, I get this with an eps=0.01

Topic 0: ; 
Topic 1: pope; 0.013

I believe that Topic 0 can be interpreted as everything else or there needs to be more topics.

arr=[]
for n in (range(2,50)):
  model = lda.LDA(n_topics=n, n_iter=20, random_state=1)
  model.fit(X)
  topic_word = model.topic_word_  # model.components_ also works
  arr.append(
    (min([max(topic_word[i]) for i in range(model.n_topics)]),
     max([max(topic_word[i]) for i in range(model.n_topics)])))
plt.plot(arr)
...

So looking at this chart, n is too low if below 5 and flattens out sometime after 20…

Notice the top word has a probability less than the epsilon when the number of topics is less than 5.  That implies that some topics are “everything else” catch alls.

Looking at the number work selected over the threshold, one can see that the optimal  threshold with this approach approaches about 27 words.
Looking at the above, notice that after a point (150 topics), the number of words chosen levels out.

Topic Words Up & Over a Threshold

This approach chooses words until the cumulative summation is greater than cut off. In this case the desire is to choose 80% of the words in a topic.

Notice that when there are too few topics far too many words are chosen. One could argue from the above graph, that the optimal bread topics looks to be around or greater than 50.

Of course the above methods are sensitive to the parameters of N, epsilon, or threshold chosen.

Log Likelihood

Thomas L. Griffiths, Mark Steyvers (2004) suggest to maximize the above graph however it appears to not be satisfactory in this situation.

Normalization

Is it better to not to normalize the prob dist and use the original value in a cut-off?

Looking at another library gensim LdaModel, it appears that LDA most likely does not originally have the probabilities sum up to 1.0 like that and they are normalized, see below:

def show_topic(self, topicid, topn=10):
    topic = self.state.get_lambda()[topicid]
    topic = topic / topic.sum() # normalize to probability dist
    ...

Running the sample code Latent Dirichlet Allocation (LDA) with Python and calling get_lambda, one can see the lambda values are sometimes above 1.0.

ldamodel.state.get_lambda()[1]

gives:

array([ 1.48214337,  1.48168697,  0.50442377,  0.50399559,  0.50400832,
    0.5047193 ,  0.50375875,  0.50376053,  1.50224118,  0.50376574,
    0.5037527 ,  0.50377459,  0.50376621,  1.49831418,  1.49832577,
    1.49831855,  1.49831883,  1.49831596,  1.51053093,  3.49684196,
    1.49832204,  1.49832512,  0.50316907,  0.50321838,  0.50328253,
    0.50319543,  0.50317986,  0.50318815,  0.50314213,  0.5031702 ,
    1.49635267,  1.49634655])

Interestingly the output of Mahout’s cvb local is not normalized to the probability distribution as the gensim library.

{dx4019:53110.68701540677,80048:35567.41624492001,..., dx13629:0.06305282918342679,dxv269:0.05556872163986669}

For the purposes of this study, output from Mahout will be normalized.

Choosing the Word Cut-Off in Claims Analysis

Interestingly, as the number of topics is increased, the initial probability of the top word decreases and levels off to about 8% or so. A possible interpretation of this topics are being split into two but perhaps they should not be. This would suggest that the number of topics may not be higher than 12.

Using the top eight work method notice that the strength of the top eight words gets diluted as the number of topics increases.

Using the threshold method the number of codes in the topic levels out after 20 or so.

When choosing the epsilon method, for determining the number of words, the number of words chosen levels off soon around 15 when epsilon=0.01 and 7 for epsilon=0.02.

Using the epsilon=0.02 technique and looking at the num_topics=3, one can see that the first topic is largely therapeutic in nature, followed by dialysis and finally the final topic being blood test related.  Oddly, a chest x-ray was included.

eps 0.02
Topic 0
97110 0.13 Therapeutic exercises
80048 0.04 Metabolic panel total ca
99213 0.04 Office/outpatient visit est
97530 0.04 Therapeutic activities
97116 0.03 Gait training therapy
97112 0.03 Neuromuscular reeducation
80053 0.03 Comprehen metabolic panel
97140 0.03 Manual therapy 1/> regions

Topic 1
36415 0.18 Routine venipuncture
q4081 0.12 Epoetin alfa, 100 units esrd
a4657 0.12 Syringe w/wo needle
90999 0.12 Dialysis procedure
j2501 0.07 Paricalcitol
99211 0.03 Office/outpatient visit est
j1756 0.02 Iron sucrose injection
j1270 0.02 Injection, doxercalciferol

Topic 2
85025 0.09 Complete cbc w/auto diff wbc
85610 0.06 Prothrombin time
80053 0.04 Comprehen metabolic panel
80061 0.03 Lipid panel
84443 0.02 Assay thyroid stim hormone
71020 0.02 Chest x-ray 2vw frontal&latl
99212 0.02 Office/outpatient visit est

Going to the other extreme and looking at the data set split into 256 topics, one can notice some topics that appear to be duplicative in nature, however perhaps they are not.  A specialist schooled in the art of dialysis may recognize an important difference.

Topic  6
a4657 0.06 Syringe w/wo needle
q4081 0.05 Epoetin alfa, 100 units esrd
85610 0.04 Prothrombin time
90999 0.03 Dialysis procedure
85025 0.03 Complete cbc w/auto diff wbc
36415 0.03 Routine venipuncture
97110 0.02 Therapeutic exercises
99213 0.02 Office/outpatient visit est
Topic  7
36415 0.09 Routine venipuncture
97110 0.07 Therapeutic exercises
q4081 0.05 Epoetin alfa, 100 units esrd
85025 0.04 Complete cbc w/auto diff wbc
a4657 0.04 Syringe w/wo needle
85610 0.03 Prothrombin time
90999 0.03 Dialysis procedure
Topic  8
90999 0.06 Dialysis procedure
97110 0.05 Therapeutic exercises
a4657 0.05 Syringe w/wo needle
85610 0.05 Prothrombin time
36415 0.03 Routine venipuncture
j2501 0.03 Paricalcitol
q4081 0.02 Epoetin alfa, 100 units esrd
80061 0.02 Lipid panel

The methods utilized in this post maybe sufficient for determining broad categories of care however it appears to not be sufficient to act as a method of determining a catalog procedures.

It is interesting to note, that with the proper choice parameters, one can approximate the frequency of the actual claim lines in the claims.

Further work in this area may be warranted.

Conclusions

LDA looks effective for discerning the high-level groupings (topics) of procedures but does not do well for this data set for finer more granular groupings.  As for the topic words themselves, using the epsilon approach for topic word selection is superior to the top N words approach while being computationally simpler than the threshold approach. The parameters involved in the epsilon method and the threshold approach can be set to achieve near identical results.   In regards the membership of each word, normalization of the lambda values makes them more interpretable and it is recommended when using Mahout.

Next Steps

As an exercise for the reader, these are the topics with the diagnosis codes included in the word sets. Do these topics look cohesive? Do the topics look disjoint?

eps 0.02

Topic  0
97110 0.14 Therapeutic exercises
97530 0.05 Therapeutic activities
97112 0.04 Neuromuscular reeducation
97140 0.04 Manual therapy 1/> regions
97116 0.03 Gait training therapy

Topic  1
36415 0.07 Routine venipuncture
85025 0.03 Complete cbc w/auto diff wbc
dx4019 0.02 Hypertension NOS
80048 0.02 Metabolic panel total ca

Topic  2
36415 0.08 Routine venipuncture
85025 0.04 Complete cbc w/auto diff wbc
dx4019 0.03 Hypertension NOS
80053 0.02 Comprehen metabolic panel
85610 0.02 Prothrombin time
80048 0.02 Metabolic panel total ca

Topic  3
q4081 0.11 Epoetin alfa, 100 units esrd
a4657 0.08 Syringe w/wo needle
90999 0.05 Dialysis procedure
j2501 0.05 Paricalcitol
j1270 0.02 Injection, doxercalciferol

Topic  4
99213 0.04 Office/outpatient visit est
36415 0.04 Routine venipuncture

Topic  5
97110 0.05 Therapeutic exercises
85025 0.02 Complete cbc w/auto diff wbc
dx4019 0.02 Hypertension NOS
80048 0.02 Metabolic panel total ca

Topic  6
80053 0.05 Comprehen metabolic panel
36415 0.04 Routine venipuncture
85025 0.03 Complete cbc w/auto diff wbc
85610 0.02 Prothrombin time
dx25000 0.02 DMII wo cmp nt st uncntr
80061 0.02 Lipid panel

Topic  7
90999 0.07 Dialysis procedure
a4657 0.07 Syringe w/wo needle
q4081 0.05 Epoetin alfa, 100 units esrd
j2501 0.04 Paricalcitol
36415 0.03 Routine venipuncture
85025 0.03 Complete cbc w/auto diff wbc

A future blog post will examine K-means. That method has a benefit of keeping the topics sets completely separate.

K-Means Claims Analysis

References

Jordan Barber (n.d.) Latent Dirichlet Allocation (LDA) with Python https://rstudio-pubs-static.s3.amazonaws.com/79360_850b2a69980c4488b1db95987a24867a.html

Chris Busch (2016) “Choosing words in a topic, which cut-off for LDA topics?” http://stats.stackexchange.com/questions/199263/choosing-words-in-a-topic-which-cut-off-for-lda-topics

Thomas L. Griffiths, Mark Steyvers (2004) “Finding scientific topics” http://psiexp.ss.uci.edu/research/papers/sciencetopics.pdf

Allen B. Riddell (2014) Topic modeling with latent Dirichlet allocation in Python https://ariddell.org/lda.html

Radim Rehurek (2011) http://pydoc.net/Python/gensim/0.8.6/gensim.models.ldamodel/