Clinical Risk Groups Analysis for Chronic Hypertensive Patients in

Terms of ICD9-CM Diagnosis Codes

Javier Fern

andez-S

anchez

, Cristina Soguero-Ruiz

, Pablo de Miguel-Bohoyo

Francisco Javier Rivas-Flores

Angel G

omez-Delgado

, Francisco Javier Guti

errez-Exp

osito

and Inmaculada Mora-Jim

enez

Dept. of Signal Theory and Communications, Telematics and Computing, Universidad Rey Juan Carlos, Madrid, Spain

University Hospital of Fuenlabrada, Fuenlabrada, Madrid, Spain

University Hospital of Sureste, Arganda del Rey, Madrid, Spain

Keywords:

Hypertension, Chronic Condition, Health Status, Clinical Risk Groups, ICD9-CM Diagnosis Codes.

Abstract:

Hypertension is a chronic condition that has a considerable prevalence in the elderly. Furthermore, hyperten-

sive patients double cost of normotensive individuals. The budget reduction and the increasing concern about

the sustainability of the healthcare system have caused that improving the efﬁciency and use of resources

are a priority in developed countries. Identiﬁcation of chronic hypertensive patients, i.e., patients with high

blood pressure, can be performed by means of population classiﬁcation systems such as Clinical Risk Groups

(CRGs). CRGs classify individuals in health status categories taking both demographic and clinical infor-

mation of the encounters that individuals have with the healthcare system during a deﬁned period of time.

In this work, we determine the characteristic proﬁle and the evolution of diagnosis codes according to the

International Classiﬁcation of Diseases 9

revision, Clinical Modiﬁcation (ICD9-CM), focusing on healthy

and chronic hypertensive patients at different chronic statuses (CRG). Our data correspond to the population

associated to the University Hospital of Fuenlabrada (Madrid, Spain) during the year 2012, providing about

46000/16000 healthy/hypertensive individuals. We found that proﬁles associated to different health statuses

have different patterns in terms of ICD-9 diagnosis codes. Furthermore, a prediction method is proposed to

determine the health status of a new patient according to demographic (age and gender) and clinical (diagnosis

codes) data. We conclude that gender is the less informative characteristic, though the combination of age and

diagnosis codes have a great potential when they are non linearly combined.

1 INTRODUCTION

The recent ﬁnancial crisis has caused an important

concern among the population, both from an econom-

ical and clinical point of view. The budget reduction

in many social areas and the increasing concern about

the sustainability of the healthcare system have been

the reasons for improving the efﬁciency and use of

resources. In this sense, the health status evolution of

chronic patients is vital for the appropriate allocation

of available resources, and therefore for the sustain-

ability of the health system.

According to (MHSSE, 2011), in Spain life ex-

pectancy increased from about 73 years in 1979 to 81

years in 2008. Since aging leads people to suffer from

chronic conditions (Anderson and Horvath, 2004), it

takes up an important role about public health spend-

ing. In fact, more than 80% of the expenditure of the

healthcare system is related to chronic diseases, and

it is estimated that 75% of the population will suf-

fer from a chronic condition anytime in their lives.

One of the chronic conditions with more prevalence

is arterial hypertension (i.e., blood pressure greater

than 140/90 mmHg), representing 18% of population

who suffer from chronic diseases according to the

Spanish National Statistical Ofﬁce. More precisely,

nearly 62% of the population aged 65 and older have

at least one chronic condition, and there is a 25%

of general population with prevalence of high blood

pressure (HBP) (Soni and Mitchell, 2016). From a

clinical viewpoint, the importance of controlling ar-

terial hypertension is high, since HBP can be asso-

ciated to the onset of other chronic clinical condi-

tions such as chronic kidney disease. It is also known

Fernández-Sánchez, J., Soguero-Ruiz, C., Miguel-Bohoyo, P., Rivas-Flores, F., Gómez-Delgado, Á., Gutiérrez-Expósito, F. and Mora-Jiménez, I.

Clinical Risk Groups Analysis for Chronic Hypertensive Patients in Terms of ICD9-CM Diagnosis Codes.

DOI: 10.5220/0006218700130022

In Proceedings of the 4th International Conference on Physiological Computing Systems (PhyCS 2017), pages 13-22

ISBN: 978-989-758-268-4

that HBP is also related to diabetes, and their com-

bination is known to have dreadful outcomes (Man-

cia et al., 2007). From an economical point of view,

the average annual expense per hypertensive patient is

about $776, with high costs in dispensation of medi-

cal drugs (Soni and Mitchell, 2016).

Nowadays, developed countries use Patient Clas-

siﬁcation Systems (PCSs) as a basic tool for cost man-

agement (Davis and LaCour, 2016), (Marazzi et al.,

2007). PCSs that take into account all patient encoun-

ters with the health system during a period of time are

named Population Classiﬁcation Systems, and they

are very useful to provide information about the pop-

ulation morbidity. Population groupers identify pa-

tients or groups of patients at high resource consump-

tion, and can be used as a mean to allocate the nec-

essary resources in health centers. There are several

population groupers, among them: Adjusted Clini-

cal Groups, mainly used for reimbursement in pri-

mary care, Diagnosis Cost Groups, and Clinical Risk

Groups (CRGs).

After analyzing different groupers, the University

Hospital of Fuenlabrada (Madrid, Spain) chose CRGs

as the most suitable system for patient classiﬁcation

(Hughes et al., 2004), (Berlinguet et al., 2005). Since

CRGs are also the most extensively used grouper, we

will consider it for research purposes. CRGs clas-

sify individuals in mutually excluding categories (risk

groups) taking demographic characteristics and clini-

cal data during a deﬁned period of time (usually, one

year). Clinical data can be provided by services of pri-

mary care, specialized care, mental health and phar-

macy dispensation. Since different codiﬁcation sys-

tems of diagnosis can be available according to the

service, a previous stage of code standardization has

to be performed by a clinical coding expert. The In-

ternational Classiﬁcation of Diseases - 9

revision-

Clinical Modiﬁcation (ICD9-CM) - has been consid-

ered in this work for diagnosis coding in primary and

specialized care.

The aim of this work is twofold. On the one hand,

to provide an explanatory analysis of both demo-

graphic data and ICD9-CM diagnosis codes related

to healthy and chronic hypertensive patients. On the

other hand, to apply statistical learning tools to pre-

dict the health status of a patient. In the long term, the

goal is to analyze clinical registers for a long enough

period to ﬁnd patterns identifying potential patients

related to chronic conditions highly prevalent in the

population. The creation of predictive models could

help to improve the health status and quality of life, as

well as to reduce the socioeconomic impact of chronic

conditions. In this sense, the work presented here can

be considered as one of the ﬁrst stages, since just one

year has been considered.

The rest of the paper is organized as follows. Sec-

tion 2 describes the HBP condition and its importance

from a clinical point of view. The PCSs are presented

in Section 3, and the database in Section 4. Section 5

presents a descriptive population analysis of ICD9-

MC Diagnosis Codes and their evolution for differ-

ent health status. A predictive analysis considering

ICD9-MC codes and demographic data is provided in

Section 6. Finally, conclusions are presented in Sec-

tion 7.

2 WHY CONSIDERING HBP?

Blood pressure is a measurement which forces your

blood against the blood vessel walls. If this force

is higher than certain value (high blood pressure), it

can produce serious damages in blood vessels over

the time. It also can increase the risk of having heart

attacks, strokes and others (Mancia et al., 2007).

HBP means high blood pressure. A person is con-

sidered to be at risk for high blood pressure when the

systolic blood pressure is 140 mm Hg or greater most

of the time, and when diastolic blood pressure is 90 or

greater mm Hg most of the time, i.e., 140/90 mm Hg.

This condition is considered as persistent in its ef-

fects with time, i.e., chronic. Despite this fact, HBP

can be controlled under blood pressure rates below

140/90 mm Hg (NLM, 2016). HBP can also lead to

get other disorders and therefore to other comorbidi-

ties. Because of this fact, it is interesting to analyze

the differences between healthy people and those who

suffer from HBP.

According to (NHLBI, 2016a), patients can be

hypertensive without knowing that they actually suf-

fer HPB, so they are at risk of developing compli-

cations because of this condition. This is the rea-

son why HBP is called the “silent killer”. Patients

do not notice any symptoms caused by HBP, rarely

they have headache. However, this condition leads to

some health problems: stretching of the blood ves-

sels would cause tears and scars that makes places for

cholesterol or blood to build up. It also would lead

to hemorrhagic strokes or aneurysms since stretch-

ing vessels can have spots which are very fragile and

would tend to break (AHA, 2016).

Though anyone could suffer from HBP, there are

some factors which can increase the probability of

having such condition (NHLBI, 2016b): gender, age,

ethnicity, overweight, and lifestyle habits, among oth-

ers. Regarding gender (Banegas, 2005), women who

are older than 55 are more likely to develop HBP than

men, though men younger than 55 have more risk to

PhyCS 2017 - 4th International Conference on Physiological Computing Systems

develop HBP. On the other hand, blood pressure in-

creases with age; speciﬁcally, nearly 60% of Span-

ish people older than 60 have HBP. Also important is

to remark that HBP can be associated to the onset of

other medical conditions such as chronic kidney dis-

ease, and it is also related to diabetes.

3 PATIENT CLASSIFICATION

SYSTEMS

Developed countries use PCSs for clinical purposes

and also as a basic tool for cost management. PCSs

are built from a set of clinical rules that assign each

patient to one of a limited collection of homoge-

neous groups in terms of resources use and/or health

status. The family based on DRGs (Diagnosis Re-

lated Groups) is the most used to analyze the hospital

case mix, as well as for reimbursement and ﬁnanc-

ing healthcare processes. DRGs were created in Yale

University at the end of 60’s (Averill et al., 1999) as a

system to classify each healthcare episode at a hospi-

tal, creating groups with clinical coherence and sim-

ilar use of hospital resources (Averill et al., 1999).

Note that every episode of the same patient can be

assigned to a different DRG according to the type of

inpatient medical assistance. Each DRG is a group

with an associated weight reﬂecting the consumption

of resources respect to an average DRG considered as

a comparison unit. Some countries use this weight,

along with other metrics (hospital stay, and others), to

compare hospital activity and funding them.

One limitation of DRGs is that they only consider

the hospital environment, but in the healthcare system

there is much more activity (primary care, specialized

care, mental health, pharmacy). PCSs that consider

different ﬁelds and take into account all the patient en-

counters with the health system are named Population

Classiﬁcation Systems or population groupers. As

previously indicated in Section 1, population groupers

usually consider age, gender, diagnoses and proce-

dures practiced during a certain period of time (Davis

and LaCour, 2016). Over the last years, pharmacy

dispensation has also been incorporated to the clini-

cal rules of some population groupers to have a more

complete knowledge about the health status.

Note that, since DRGs are based on individual

healthcare episodes, they do not take into account the

information of the same patient during a period of

time. Since chronic conditions tend to be reﬂected in

certain diagnosis and pharmacy dispensation, DRGs

are not the most appropriate system to identify pa-

tients with chronic conditions. A classiﬁcation sys-

tem considers all encounters with the health system

is necessary to identify patients with chronic condi-

tions. Population groupers named CRGs (developed

by 3M

) take into account data about the same pa-

tient for a period of time (usually one year (Newhouse

et al., 1997)) and assign the patient to a single mutu-

ally exclusive risk group which relates the clinical and

demographic characteristics of every individual. This

way, CRGs are suitable for identifying patients with

different chronic conditions and severity levels. This

will allow to relate an individual to the amount and

type of healthcare resources that the patient will need

in the future (Averill et al., 1999).

By way of a summary, the main differences be-

tween DRGs and CRGs are the following: (1) CRGs

categorize individuals whereas DRGs classify inpa-

tient medical assistance; (2) DRGs classify the pa-

tient into an inpatient service episode considering data

registered during the stay, whilst CRGs may consider

hospital environment in addition to primary care, out-

patient care, mental health, and pharmacy (Averill

et al., 1999).

Regarding CRGs, there are 9 core health status

groups, which are subdivided into 272 chronic condi-

tion categories (ccc). The core groups are hierarchi-

cally organized as follows: CRG-9 (catastrophic, 11

ccc), CRG-8 (metastatic malignancy, 22 ccc), CRG-

7 (chronic triplet, 21 ccc), CRG-6 (multiple signif-

icant chronic pair, 61 ccc), CRG-5 (single domi-

nant/moderate chronic, 107 ccc), CRG-4 (multiple

minor chronic pair, 1 ccc), CRG-3 (single minor

chronic, 41 ccc), CRG-2 (signiﬁcant acute, 6 ccc),

and CRG-1 (healthy user and non-user, so 2 ccc). Ad-

ditionally, every chronic condition category is subdi-

vided into severity levels (up to 6, depending on the

type of condition), providing a total of 1080 groups.

A number of 5 digits is considered to code every of

the 1080 groups. First digit is associated to the core

group, next three digits refer to the ccc, and last digit

is associated to the severity level.Thus, CRG-51913

corresponds to the group of single dominant/moderate

chronic diseases (ﬁrst digit), and speciﬁcally refers to

coronary atherosclerosis (next three digits), with the

third level of severity (last digit).

Patients with chronic condition of HBP can be

classiﬁed into 34 different groups (of a total of 1080):

lower CRG number refers to lower chronic condition.

Owing to the limited number of patients associated

to some severity levels, we chose to merge groups of

different severity levels when they are related to the

same chronic condition. From a statistical point of

view, it is important the availability of groups of rea-

sonable size. This way, our analysis will be focused

on the kind of chronic condition, and groups with the

same ﬁrst four digits are merged to create what we

Clinical Risk Groups Analysis for Chronic Hypertensive Patients in Terms of ICD9-CM Diagnosis Codes

have called base-CRGs. Thus, base-CRG 5192 en-

compasses patients with HBP and four severity lev-

els (CRG 51921, CRG 51922, CRG 51923, and CRG

51924).

4 DATABASE DESCRIPTION

The implementation of the Electronic Health Record

(EHR) in the current society has become a power-

ful tool in organizational terms (annotation legibility,

content security, paper ﬁles removal, etc) (H

ayrinen

et al., 2008). In this work, data from the EHR of Uni-

versity Hospital of Fuenlabrada along the year 2012

have been analyzed. This hospital assists to several

small towns in the south area of Madrid, encompass-

ing about 225.000 inhabitants. Speciﬁcally, we con-

sider clinical encounters and demographic data (age

and gender) from individuals categorized by the CRG

system as healthy and hypertensive ones by using

gender, age, clinical diagnoses, procedures and phar-

macy dispensation.

Hypertensive patients are associated to one out of

six different base-CRG, coded by four numbers: base-

CRG 5192 (hypertension), base-CRG 6124 (chronic

obstructive pulmonary disease and hypertension),

base-CRG 6144 (diabetes and hypertension), base-

CRG 6242 (asthma and hypertension), base-CRG

7070 (diabetes, cerebrovascular disease and hyper-

tension), and base-CRG 7071 (diabetes, hypertension

and other dominant chronic disease). Just one base-

CRG related to HBP has been considered per core

group in this work. The chosen base-CRG has been

the one with more individuals: base-CRG 5192, base-

CRG 6144, and base-CRG 7071. Patients classiﬁed

in CRG-1 (healthy) have been also considered in our

study, providing a total of 63008 individuals, as de-

tailed in Table 1.

Table 1: Number of individuals in year 2012 per base-CRG.

Healthy status is associated to CRG-1.

base-CRG Individuals in 2012

Healthy 46835

5192 12447

6144 3179

7071 547

TOTAL 63008

Figure 1 shows the age distribution for hyperten-

sive and healthy patients. Note that healthy patients

are younger than hypertensive patients, what is quite

reasonable. Regarding patients with HBP, age of

individuals in the lower base-CRG (no comorbidity)

tend to be younger than individuals with comor-

bidities (base-CRG 6144 and base-CRG 7071).

Figure 1: Age histogram for individuals in CRG-1 (healthy,

upper panel) and the three base-CRGs related to HBP.

Table 2 shows the gender distribution per base-

CRG, both for healthy and hipertensive patients. Gen-

erally speaking, there is a balance between both gen-

ders except for base-CRG 7071, with the highest pro-

portion for women. This fact can be explained by an

increase of chronic conditions as aging, with a higher

life expectancy for women.

Table 2: Gender distribution (in %) for healthy and hyper-

tensive patients according to the corresponding base-CRG.

Men Women

Healthy CRG-1 47.3 52.7

base-CRG 5192 46.9 53.1

base-CRG 6144 54.0 46.0

base-CRG 7071 42.6 57.4

Next, Figure 2 and Figure 3 show the distribution

of patients with an speciﬁc number of diagnosis codes

per year and health status. Horizontal axis in both ﬁg-

ures refer to the number of diagnosis codes per pa-

tient: total number of codes (Figure 2) and number

of different codes (Figure 3). Note that healthy indi-

viduals have a lower number of diagnosis codes than

hypertensive patients, what seems quite reasonable.

Additionally, the mode in the number of different di-

agnosis codes per patient increases with the chronic

condition, going from 1 in the healthy group to 6 in

the base-CRG 7071. Note also that the distribution

tends to be heavy-tailed in the right hand side as the

chronic condition worsens. These results are consis-

tent with the comorbidity associated to every base-

CRG.

PhyCS 2017 - 4th International Conference on Physiological Computing Systems

Figure 2: Histogram of the total number of diagnosis codes

per patient and health status.

Figure 3: Histogram of the number of different diagnosis

codes per patient and health status.

5 POPULATION ANALYSIS OF

ICD9-MC DIAGNOSIS CODES

A population analysis to get the average proﬁle of

ICD9-MC codes and the most frequent codes per

base-CRG is ﬁrst presented in Section 5.1. Then, the

evolution of codes associated to health statuses is an-

alyzed in Section 5.2.

5.1 Average ICD9-CM Diagnosis Code

Proﬁles

The syntax of ICD9-CM codes is based on categories,

subcategories and sub-classiﬁcations of diseases. As

such, the three ﬁrst digits of the code refer to the cat-

egory. This is followed by a dot and then the sub-

category is represented by the next digit. The second

digit after the dot refers to the sub-classiﬁcation. For

instance, diagnoses codes (000-999) can be written

as “XXX.XX”, V codes (V01-V89) as “VXX.XX”, E

codes (E800-E999) as “EXXX.X”, M codes (M8000-

M9970), and procedures (00-99) as “XX.XX”, where

X is a digit number between 0 and 9. For simplicity,

from now on we will just consider codes referring to

categories (three digits/values for diagnosis codes).

We deﬁne the average proﬁle as a bi-dimensional

graph representing the presence rate of every code for

patients associated to a speciﬁc group (base-CRG).

Thus, the average proﬁle is obtained taking into

account the presence/absence of diagnosis codes per

patient in a group. Then, the presence rate per code is

computed, resulting in a number in the interval [0,1].

This number represents the rate of patients having

a speciﬁc diagnosis code. We focus on individuals

from CRG-1 (healthy users), base-CRG 5192 (HBP),

base-CRG 6144 (HBP and diabetes) and base-CRG

7071 (HBP, diabetes and other dominant chronic

condition). For each group, we aim to show if there

is a pattern associated to ICD9-MC diagnosis codes.

Average Diagnosis Code Proﬁle for the Healthy

Status Group. Figure 4 (a) shows the average diag-

nosis proﬁle of ICD9-MC diagnosis codes for individ-

uals classiﬁed as healthy. The horizontal axis repre-

sents the ICD9-MC codes: diseases codes (from 1 to

1000), V codes (from 1001 to 1091), E codes (1092 to

1319) and M codes (1320 to 1517). The vertical axis

represents the rate of patients having a speciﬁc diag-

nosis code. For more detail, see Figure 5 (a), where

the 15 diagnosis codes with highest rate are displayed.

According to this ﬁgure, there is no prevailing

code, with the highest rate lower than 15% for healthy

patients. The highest rate correspond to jaws issues

(526) at the ﬁrst place, along with common cold (460)

and general symptoms (780). Other diagnoses with

prevalence are acute tonsillitis (463) and other issues

related with gastroenteritis (558).

These results seem reasonable, since the most

common diagnoses are acute diseases and they can

be considered as normal issues in healthy individuals.

Clinicians validated that these codes are related to

common diseases.

Average Diagnosis Code Proﬁle for base-CRG

5192. Figure 4 (b) presents the presence rate of diag-

nosis codes when considering patients assigned to the

base-CRG 5192. As expected, the most common di-

agnosis is HBP (401), which dominates considerably

above all; followed by excess of low-density lipopro-

tein (LDL) Cholesterol (272) in to a lesser extent. In

fact, HBP and high LDL cholesterol (also known as

“bad” cholesterol) are two of the most common, seri-

ous and treatable medical conditions that lead to car-

diovascular disease (Parks et al., 2006). The average

proﬁle also shows that Back issues (724), joint issues

such as hermarthrosis or synovitis (719) or general

Clinical Risk Groups Analysis for Chronic Hypertensive Patients in Terms of ICD9-CM Diagnosis Codes

(a)

(b)

(c)

(d)

Figure 4: Average proﬁle of ICD9-MC diagnosis codes of

individuals classiﬁed into: (a) CRG-1 (healthy); (b) base-

CRG 5192; (c) base-CRG 6144, and (d) base-CRG 7071.

symptoms (780) are also relevant. For more informa-

tion, see Fig. 5 (b), where codes associated to 15 high-

est rates are displayed.

Prevalence of the 401 ICD9-MC code (essential

hypertension) for patients assigned to the base-CRG

5192 is a key characteristic on this group. As a

difference from the average proﬁle of CRG-1, the

highest rate is about 80% for hypertensive patients.

Note that the PCS also consider medical drugs to

assign a patient to a CRG, so it is possible that the

remaining 20% of patients are assigned to base-CRG

5192 because they take hypertensive drugs.

Average Diagnosis Code Proﬁle for base-CRG

6144. As we expected, Diabetes mellitus (DM) (250)

and HBP (401) take relevance above other codes in

this CRG (see Figure 4 (c) and Figure 5 (c) for de-

tails), with rates about 80%, which is consistent with

the CRG analyzed. In a lesser extent, the following

rates are related to excess of LDL cholesterol (272),

joint issues (719) and back issues (724). Diagnosis

codes related to common cold (460) and general

symptoms (780) appear in a similar rate to that of

healthy group.

Average Diagnosis Code Proﬁle for base-CRG

7071. As in base-CRG 6144, diagnosis codes related

(a)

(b)

(c)

(d)

Figure 5: Presence rate for the 15 diagnosis codes with

highest rate in: (a) CRG-1 (healthy); (b) base-CRG 5192;

to DM (250) and HBP (401) are those with a high-

est rate (see Figure 4 (d)). Note that nearly 80% of

patients in this base-CRG were assigned at least one

of these codes, a similar proportion as in base-CRG

6144. An important difference respect to the proﬁle

of base-CRG 6144 is that codes related to excess of

LDL cholesterol (272), and chronic obstruction of the

airways to the lungs (496) are also predominate (see

Figure 5 (d)). This shows that in this CRG more co-

morbidities are present, being the patient in a worst

chronic condition.

The average proﬁle allows us to validate clini-

cal knowledge, since DM, HBP, and another chronic

dominant condition are the diagnosis codes with most

presence in each patient.

As a summary, Table 3 lists the 5 most proba-

ble diagnosis codes per CRG. Comparing the healthy

status group with other CRGs, we can claim that

healthy patients do not get a high percentage of com-

mon diseases among them, which means that diseases

within the healthy status group have more variabil-

ity. Because of that, the most probable disease has

a prevalence lower than 15% of the healthy individu-

als, which is jaws issues (14.8%). For the other ana-

lyzed CRGs, we can conﬁrm that there is a high prob-

ability to have chronic diseases, being HBP (74.9%

in base-CRG 5192, 71.73% in base-CRG 6144 and

PhyCS 2017 - 4th International Conference on Physiological Computing Systems

Table 3: More frequent ICD9-MC diagnosis codes ex-

pressed in % in individuals classiﬁed into the healthy status

group, CRG 5192, CRG 6144 and CRG 7071.

Code Description %

526 Jaws issues 14.8

460 Common cold 12.3

Healthy 780 General symptoms 6.1

463 Acute tonsillitis 8.8

558

Other issues related

with gastroenteritis

8.6

401 Essential HBP 74.9

272

Excess of LDL

(cholesterol)

20.3

CRG 5192 724 Back issues 14.0

719 Joint issues 12.9

780 General symptoms 10.5

250 DM 87.4

401 Essential HBP 71.3

CRG 6144 272

Excess of LDL

(cholesterol)

27.3

724 Back issues 13.8

719 Joint issues 17.1

250 DM 86.8

401 Essential HBP 69.2

780 General symptoms 31.1

CRG 7071 272

Excess of LDL

(cholesterol)

28.7

496

Chronic obstruction

of the airways

to the lungs

20.1

69.2% in base-CRG 7071) and DM (87.4% in CRG

6144 and 86.8% in CRG 7071) the most probable dis-

eases found within those groups.

5.2 Evolution of ICD9-CM Codes in

Terms of Health Status

Figure 6 shows the evolution of the diagnosis codes

with higher presence rate according to the health sta-

tus (CRG) in 2012. Note that nearly 75% of the hyper-

tensive patients have the diagnosis code 401 (HBP),

while DM is present in more than 85% of patients as-

signed to base-CRG 6144 and base-CRG 7071. It is

clear from Figure 6 how excess of LDL (cholesterol,

diagnosis code 272) increases its prevalence with the

number of the health core status (from groups 5192

to 7071). It is in base-CRG 7071, associated to the

worst chronic condition among those considered here,

where excess of LDL takes the highest presence rate

(about 30%). Code with the highest presence rate in

the group of healthy patients (jaws issues, 526) have

a similar rate in all base-CRGs (presence rate lower

than 15%). General symptoms (780) is also present

in every group, having the highest rate in base-CRG

7071 (about 30%), nearly tripling its presence regard-

ing healthy patients. This fact can be justiﬁed by the

fact that patients in the seventh core group have a high

number of comorbidities and so the general symptoms

code is quite frequent when attending to the doctor.

We conclude stating that the evolution of codes over

different CRGs was as expected prior the analyses:

HBP presents a high margin over acute diseases in

base-CRGs 5192, 6144 and 7071, getting a similar

behaviour for DM in the last two CRGs. Besides, we

verify the relationship between healthy patients and

acute diseases, these being present in a low rate.

Figure 6: Evolution of the ICD9-CM diagnosis codes with

the highest presence rate in terms of health status (CRG).

6 HEALTH STATUS PREDICTION

The population analysis of ICD9-MC diagnosis codes

supports that patients classiﬁed according to different

health status (CRG) are characterized by a different

diagnosis code proﬁle, as presented in Fig. 5. This

suggests that a predictive model can be built to pre-

dict the base-CRG of a new patient. Towards that end,

we evaluate linear and nonlinear models constructed

by using a Support Vector Machine (SVM). The se-

lection of this method is motivated by its theoretical

properties, which make it an attractive approach for

a great number of medical data problems (Soguero-

Ruiz and et al., 2016a), (Soguero-Ruiz and et al.,

2016b).

Both linear and nonlinear SVM classiﬁers require

free parameters tuning (Steinwart and Christmann,

2008). The linear ν-SVM algorithm requires the tun-

ing of a single free parameter ν ∈ (0, 1), while the

spread parameter (σ) has to be also adjusted for the

Clinical Risk Groups Analysis for Chronic Hypertensive Patients in Terms of ICD9-CM Diagnosis Codes

Figure 7: Average presence rate for the 200 diagnosis codes

with the higher values.

nonlinear SVM when a Gaussian kernel is considered.

We adopted a cross-validation strategy for tuning the

free parameters.

In this work, demographic (gender and age) and

clinical data (ICD9-CM diagnosis codes) have been

considered as input features to classify a patient into

one of the following status groups: CRG-1, base-

CRG 5192, base-CRG 6144 or base-CRG 7071.

Since classes are highly unbalanced (see Table 1),

the data set was constructed using an undersampling

strategy in order to get balanced classes. Towards that

end, a random subset of 547 samples were selected for

each group, achieving a total of 2188 samples when

considering the four groups. Note that the number of

samples per group is limited by the minority group,

i.e. base-CRG 7071. On the balanced data set, 80%

of the samples were randomly chosen for training, be-

ing the remaining 20% for testing. We repeated this

process 50 times and evaluated results in terms of ac-

curacy (mean ± standard deviation).

We have checked that when all features (demo-

graphic and diagnosis codes) were considered as in-

put features in the SVM, the model accuracy was

low. Therefore, results are not presented in this paper.

This low accuracy can be motivated because there are

many diagnosis codes with a very low presence rate

in all groups (see Fig. 4). To tackle this shortcoming,

we averaged the proﬁles presented in Section 5.1 to

get an average presence rate, regardless the number

of patients per group. Figure 7 shows the average rate

of the 200 diagnosis codes with the highest average

presence rate.

To enhance the prediction results, and according

to Figure 7, the 50 diagnosis codes with higher aver-

age presence rate were evaluated (see Figure 8 for de-

tails). Using these diagnosis codes, a binary classiﬁ-

cation was ﬁtted for: (1) CRG-1 vs. base-CRG 5192;

(2) CRG-1 vs. base-CRG 6144; (3) CRG-1 vs. base-

CRG 7071; (4) base-CRG 5192 vs. base-CRG 6144;

(5) base-CRG 5192 vs. base-CRG 7071; and (6) base-

CRG 6144 vs. base-CRG 7071. In general, the best

performance (mean accuracy of 97%) was achieved

when classifying patients from CRG-1 vs. base-CRG

7071. This is an expected result since healthy pa-

tients are characterized by having different diagnosis

codes with respect to chronic triplet ones (see proﬁles

in Figure 4). Accuracy decreases signiﬁcantly when

just comparing patients with chronic conditions. For

example, a mean accuracy value of 76% is obtained

when classifying patients with chronic triplet (base-

CRG 7071) from patients with a chronic pair (base-

CRG 6144).

On the way to predict the health status evolution,

a multiclass SVM was trained, evaluating four classes

(one for each group in Table 1). Table 4 shows the ac-

curacy (mean and standard deviation) when using de-

mographic and clinical data as input features, both in-

dividually and jointly for linear and non linear SVM.

Results suggest that a non linear approach provides

better performance than a linear one. On the other

hand, clinical features provide more knowledge than

demographic data (ﬁrst and second rows in Table 4).

Note that accuracy is improved when age and clinical

features are jointly analyzed, with gender being the

less informative feature.

7 CONCLUSIONS

Nowadays, HBP plays an important role in a clini-

cal and economic context, with a high incidence and

prevalence among mid-aged and older people (Soni

and Mitchell, 2016). Besides, treatment and control

of this particular disease can be expensive (Soni and

Mitchell, 2016).

This work aims to provide knowledge which can

support clinicians decisions as well as improve the

resources allocation devoted to the speciﬁc chronic

condition of HBP. For that purpose, we analyzed the

evolution of ICD9-CM diagnosis codes, focusing on

healthy and hypertensive patients at different chronic

status. Furthermore, we predict the base-CRG for

a new patient. To that end, we analyzed data from

EHRs of UHF. These data refer to ICD9-MC diag-

noses codes, along with demographic variables asso-

ciated to 63.008 individuals in the year 2012.

Firstly, an exploratory analysis of these individu-

als was performed. There is a fair similar number of

individuals of each gender in every CRG. We verify

that women’s life expectancy is higher than that of

men by taking gender distribution of different hyper-

PhyCS 2017 - 4th International Conference on Physiological Computing Systems

Figure 8: Average presence rate and ICD9-CM description for the 50 codes with higher values.

Table 4: Accuracy values considering the 50 diagnosis codes with highest presence rates. Both linear and non linear SVM

have been considered for classifying patients into one of four health status (CRG-1, base-CRG 5192, base-CRG 6144 and

base-CRG 7071).

Feature Linear Non linear

Demographic Gender 0.25 ± 0.02 0.26 ± 0.02

Age 0.35 ± 0.10 0.49 ± 0.02

Gender & Age 0.32 ± 0.10 0.50 ± 0.02

Clinical CIE9-CM codes 0.55 ± 0.05 0.75 ± 0.01

Demographic Gender & CIE9-CM 0.55 ± 0.06 0.75 ± 0.01

and Age & CIE9-CM 0.62 ± 0.04 0.76 ± 0.02

Clinical Age & Gender & CIE9-CM 0.63 ± 0.04 0.76 ± 0.02

tensive groups into account. We also check that the

older the patients, the more comorbidities they have

(Figure 1 shows that age increases from base-CRG

5192 until base-CRG 7071).

We also analyzed the number of diagnosis codes

per patient and health status by means of a population

analysis, claiming that individuals belonging to CRG-

1 have much less diagnosis codes per patient than hy-

pertensive ones. According to the average ICD9-CM

diagnosis code proﬁle for CRG-1, the most probable

code has a prevalence rate lower than 15%, whereas

the most probable code in the hypertensive-related

CRGs is around 70-80%. The most relevant codes

in the studied chronic groups are HBP and DM, as

expected prior to perform this work.

Furthermore, a health status prediction method is

proposed to predict the base-CRG of a new patient.

Speciﬁcally, linear and non linear SVM classiﬁers

are trained to evaluate demographic and clinical data

both individually and jointly. The highest accuracy

is obtained when a nonlinear SVM is run in a sub-

set of ICD9-CM codes with age and gender. These

results are promising, since the considered gold stan-

dard (health status group) has been achieved by also

feeding the PCS with medical drug data, not stud-

ied in this paper. As a future work, we will include

also information about medical drugs, which will en-

hance classiﬁer performance and report more knowl-

edge about chronic condition in modern societies.

In conclusion, our analysis has allowed us to de-

scribe and verify the relationship that demographic

and standardized clinical data, collected from patient

encounters with primary and specialized care, have

with the patient health status.

ACKNOWLEDGEMENTS

This work is partially supported by Spanish Research

Project TEC2016-75361-R. Cristina Soguero-Ruiz is

supported by FPU grant AP2012-4225 from Spanish

Government. The authors would like to thank the staff

of Healthcare Iberia 3M for their support and assis-

tance with this project.

REFERENCES

AHA (2016). American health association.

http://www.heart.org [Accessed: 27/05/2016].

Clinical Risk Groups Analysis for Chronic Hypertensive Patients in Terms of ICD9-CM Diagnosis Codes

Anderson, G. and Horvath, J. (2004). The growing burden

of chronic disease in america. Public health reports,

119(3):263–270.

Averill, R. F., Goldﬁeld, N., Eisenhandler, J., Hughes, J.,

Shaﬁr, B., Gannon, D., Gregg, L., Bagadia, F., Stein-

beck, B., Ranade, N., et al. (1999). Development and

evaluation of clinical risk groups (crgs). Wallingford,

CT: 3M Health Information Systems.

Banegas, J. B. (2005). Epidemiology of arterial hyper-

tension in spain. present situation and perspectives.

Hipertensi

on, 22(9):353–362.

Berlinguet, M., Preyra, C., and Dean, S. (2005). Com-

paring the value of three main diagnostic-based risk-

adjustment systems (dbras). Technical report, Cana-

dian Health Services Research.

Davis, N. A. and LaCour, M. (2016). Foundations of Health

Information Management. Elsevier Health Sciences.

ayrinen, K., Saranto, K., and Nyk

anen, P. (2008). Def-

inition, structure, content, use and impacts of elec-

tronic health records: a review of the research liter-

ature. International journal of medical informatics,

77(5):291–304.

Hughes, J. S., Averill, R. F., Eisenhandler, J., Goldﬁeld,

N. I., Muldoon, J., Neff, J. M., and Gay, J. C. (2004).

Clinical risk groups (crgs): a classiﬁcation system

for risk-adjusted capitation-based payment and health

care management. Medical care, 42(1):81–90.

Mancia, G., De Backer, G., Dominiczak, A., Cifkova, R.,

Fagard, R., Germano, G., Grassi, G., Heagerty, A. M.,

Kjeldsen, S. E., Laurent, S., et al. (2007). 2007 guide-

lines for the management of arterial hypertension. Eu-

ropean heart journal, 28(12):1462–1536.

Marazzi, A., L, G., and HD, D. (2007). New approaches

to reimbursement schemes based on patient classiﬁca-

tion systems and their comparison. Health Serv Man-

age Res, 20(3):203–210.

MHSSE (2011). National health system annual report.

Technical report, Ministry of Health Social Services

and Equality. Sistema Nacional de Salud.

Newhouse, J. P., Buntin, M. B., and Chapman, J. D. (1997).

Risk adjustment and medicare: taking a closer look.

Health Affairs, 16(5):26–43.

NHLBI (2016a). National Heart, Lung, and Blood Institute.

http://www.nhlbi.nih.gov/ [Accessed: 27/05/2016].

NHLBI (2016b). Risk Factors for High Blood

Pressure. http://www.nhlbi.nih.gov/health/health-

topics/topics/hbp/atrisk [Accessed: 29/05/2016].

NLM (2016). National library of medicine. high blood pres-

sure. www.goo.gl/Lkhkkc [Accessed: 27/05/2016].

Parks, J., Svendsen, D., Singer, P., Foti, M. E., and Mauer,

B. (2006). Morbidity and mortality in people with

serious mental illness. Alexandria, VA: National As-

sociation of State Mental Health Program Directors

(NASMHPD) Medical Directors Council, 25.

Soguero-Ruiz, C. and et al. (2016a). Predicting colorec-

tal surgical complications using heterogeneous clini-

cal data and kernel methods. Journal of Biomedical

Informatics, 61:87 – 96.

Soguero-Ruiz, C. and et al. (2016b). Support vector fea-

ture selection for early detection of anastomosis leak-

age from bag-of-words in electronic health records.

IEEE Journal of Biomedical and Health Informatics,

20(5):1404–1415.

Soni, A. and Mitchell, E. (2016). Expenditures for com-

monly treated conditions among adults age 18 and

older in the u.s. civilian noninstitutionalized popula-

tion, 2016. Statistical Brief, 487. May 2016. Agency

for Healthcare Research and Quality, Rockville, MD

[Accessed: 22/02/2017].

Steinwart, I. and Christmann, A. (2008). Support vector

machines. Springer Science & Business Media.

PhyCS 2017 - 4th International Conference on Physiological Computing Systems