Content uploaded by Glenn Flores
Author content
All content in this area was uploaded by Glenn Flores on Jan 28, 2022
Content may be subject to copyright.
Comparing Artificial Intelligence and Traditional
Methods to Identify Factors Associated With
Pediatric Asthma Readmission
Alexander H. Hogan, MD, MS; Michael Brimacombe, PhD; Maua Mosha, MS;
Glenn Flores, MD
From the Division of Hospital Medicine, Connecticut Children’s Medical Center (AH Hogan), Hartford, Conn; Department of Pediatrics,
University of Connecticut School of Medicine (AH Hogan), Farmington, Conn; Health Services Research Institute, Connecticut Children’s
Medical Center (M Brimacombe and M Mosha), Hartford, Conn; and Department of Pediatrics, University of Miami Miller School of
Medicine, and Holtz Children’s Hospital, Jackson Health System (G Flores), Miami, Fla
The authors have no conflicts of interest to disclose.
Address correspondence to Alexander H. Hogan, MD, MS, Connecticut Children’s Medical Center, 282 Washington St, Hartford, CT 06106
(e-mail: AHogan@connecticutchildrens.org).
Received for publication December 30, 2020; accepted July 20, 2021.
TAGGEDPABSTRACT
OBJECTIVE:To identify and contrast risk factors for six-month
pediatric asthma readmissions using traditional models (Cox
proportional-hazards and logistic regression) and artificial neu-
ral-network modeling.
METHODS:This retrospective cohort study of the 2013 Nation-
wide Readmissions Database included children 5 to 18 years
old with a primary diagnosis of asthma. The primary outcome
was time to asthma readmission in the Cox model, and read-
mission within 180 days in logistic regression. A basic neural
network construction with 2 hidden layers and multiple repli-
cations considered all dataset variables and potential variable
interactions to predict 180-day readmissions. Logistic regres-
sion and neural-network models were compared on area-
under-the receiver-operating curve.
RESULTS:Of 18,489 pediatric asthma hospitalizations, 1858
were readmitted within 180 days. In Cox and logistic models,
longer index length of stay, public insurance, and nonwinter
index admission seasons were associated with readmission
risk, whereas micropolitan county was protective. In neural-
network modeling, 9 factors were significantly associated with
readmissions. Four overlapped with the Cox model (nonwin-
ter-month admission, long length of stay, public insurance,
and micropolitan hospitals), whereas 5 were unique (age, hos-
pital bed number, teaching-hospital status, weekend index
admission, and complex chronic conditions). The area under
the curve was 0.592 for logistic regression and 0.637 for the
neural network.
CONCLUSIONS:Different methods can produce different read-
mission models. Relying on traditional modeling alone over-
looks key readmission risk factors and complex factor
interactions identified by neural networks.
TAGGEDPKEYWORDS:artificial neural network; asthma; machine learn-
ing; rehospitalization
ACADEMIC PEDIATRICS 2022;22:55−61
TAGGEDPWHAT’S NEW
An artificial neural network performed slightly better
than traditional statistical models, and identified risk
factors not identified by traditional modeling. Key
readmission risk factors common to both models were
non-winter-month admission, long length of stay, pub-
lic insurance, and micropolitan hospitals.
TAGGEDPASTHMA IS ONE of the most common chronic conditions
of childhood in the United States. Although only 1% to
3% of children with asthma are hospitalized each year,
hospitalizations account for over one third of asthma
spending nationally, and create significant financial bur-
dens for patients, families, and the health care system.
1,2
Early asthma readmissions (<30 days) are both infrequent
(1%−2%) and unlikely to be preventable: one pediatric
study found that the all-cause preventable early-readmis-
sion rate was <2%.
3
Late readmissions are more common,
and thus a larger burden on families—4% to 10% of chil-
dren are readmitted within 180 days
4,5
This is concerning,
as asthma is an ambulatory-care-sensitive condition for
which timely, appropriate outpatient care could prevent
hospitalization.
6
Multicomponent quality-improvement
interventions have decreased late readmissions, but none
has used readmission risk assessments to identify and tar-
get those at highest risk of readmission.
7
Establishing the
best modeling techniques and most salient readmission
risk factors may aid in future readmission reduction
endeavors.
Readmission modeling has traditionally employed Cox
and logistic regression. Readmission prediction models
using traditional methods for analyzing administrative
datasets have yielded less-than-optimal results, with an
area-under-the-receiver-operator curve (AUC) of 0.55 to
0.65.
8
These findings may be due to a lack of granular
data, or there may be limitations in analysis techniques.
Machine-learning approaches, specifically artificial neural
ACADEMIC PEDIATRICS
Copyright ©2021 by Academic Pediatric Association 55
Volume 22, Number 1
January−February 2022
networks (ANNs), have been proposed as potentially
superior predictive modeling strategies. The key advan-
tages of ANNs are the intrinsic evaluation of nonlinear
relationships that are difficult to achieve in traditional sta-
tistical models. The assessment of variable interplay goes
beyond the standard multiplicative assessment of variable
interaction of traditional statistical models to incorporate
variable correlations and interactions on a scale beyond
what is possible in a traditional regression model. The
optimal method for predicting pediatric asthma readmis-
sions and identifying readmission risk factors at 180 days
is unknown. The study objective was to use traditional sta-
tistical modeling (Cox proportional-hazards and logistic
regression) and ANN models to identify and contrast risk
factors and model performance for 180-day pediatric
asthma readmissions in a nationally representative sam-
ple.
TAGGEDH1METHODSTAGGEDEND
TAGGEDH2STUDY DESIGN AND SETTINGTAGGEDEND
This was a 1-year retrospective cohort analysis of the
2013 Agency for Healthcare Research and Quality
Nationwide Readmission Database (NRD), an all-payer
claims database with linkages across hospitalizations.
9
The NRD consists of 49% of US inpatient and observation
hospitalizations, and is drawn from 22 state inpatient data-
bases. The unit of analysis for the NRD is discharge
records. De-identified patient-level discharge records,
with verified individual identifiers, track patients across
hospitals within a state. No attempts were made to identify
any individuals in the database. All parties with access to
NRD adhered to Healthcare Cost and Utilization Project’s
formal data use agreement. The study was deemed exempt
from full review by Connecticut Children’s Medical Cen-
ter institutional review board.
We examined the question that researchers, payors, and
health systems are most likely to ask: what are the key
predictors of 180-day readmissions in a given calendar
year, using different analytic methods?
T
AGGEDH2STUDY POPULATIONTAGGEDEND
The study sample included hospitalizations of children
5 to 18 years old who had an index admission for asthma,
as defined by a primary diagnosis code for asthma in the
International Classification of Disease, Ninth Revision,
Clinical Modification (ICD-9 code: 493.xx), and were
admitted between January 1, 2013, and December 31,
2013. We excluded discharges of patients who transferred
to another acute hospital, left against medical advice, or
died during the hospital stay. Discharged patients <5 years
old also were excluded to minimize the number of admis-
sions due to bronchiolitis and virus-induced wheeze.
T
AGGEDH2PRIMARY OUTCOMETAGGEDEND
The primary outcome was whether or not a child had an
asthma readmission within 180 days. Similar to prior stud-
ies, this outcome was defined as the first hospitalization a
primary diagnosis of asthma (ICD-9 code = 493.xx) fol-
lowing discharge from the index hospitalization.
10
TAGGEDH2INDEPENDENT VARIABLESTAGGEDEND
Sociodemographic variables included sex, age (dichot-
omized as 5−11 and 12−18 years old to correspond to
age groups for the National Heart, Lung, and Blood Insti-
tute asthma guidelines
11
), residential income quartile,
9
insurance payer (private, public, uninsured, and other),
social risk factors (present or absent, based on social
determinants of health ICD-9 codes for family member
with alcohol and/or drug problem, history of abuse, paren-
tal separation, foster care, educational circumstance,
housing instability, other economic strain, and legal cir-
cumstance),
12
and patient county residence, classified by
population density as large metropolitan (population >1
million), small metropolitan (50,001−1,000,000), micro-
politan (10,001−50,000), or nonurban/rural (≤10,000).
13
Race/ethnicity is not included in the NRD, and so could
not be assessed. Patient clinical characteristics included
presence of complex chronic conditions (CCCs),
14
and
length of stay (LOS), categorized as short (<2 days),
median (2−3 days) and long (≥4 days, the top quartile of
LOS), admission season (spring [March, April, May],
summer [June, July, August], fall [September, October,
November], or winter [December, January, February]),
and disposition at discharge (home, home with nursing
services, postacute care, or unknown).
9
Due to sampling
practices and privacy measures, patients identifiers are not
tracked between years in the NRD; therefore, prior-year
admissions were not included. Hospital characteristics
assessed included hospital ownership (government, pri-
vate not-for-profit, or private for-profit), teaching status,
and hospital size by number of beds (small, medium, or
large stratified by US region and rural or urban hospital
location).
15
Markers of hospitalization severity, such as
All Patient Refined Diagnosis Related Group severity and
the Hospitalization Resource Intensity Score for Kids,
cannot be determined until after discharge, when billing
and coding are completed. As such, these scores were not
considered for adjustment in the analysis.
T
AGGEDH2STATISTICAL ANALYSISTAGGEDEND
Discharges with and without a readmission were com-
pared using chi-square tests for all categorical variables.
The Wilcoxon signed rank test was used to compare con-
tinuous variables. Subjects with missing data were
excluded from analysis. Two methods of traditional statis-
tical model construction were used: Cox proportional haz-
ards and logistic regression. The primary analysis was
Cox proportional hazards, as it was best able to account
for patient time to readmission event in the study. As the
primary outcome of readmission within 180 days neces-
sarily censored patients admitted in the second half of the
year, the Cox proportional-hazards model best accounted
for patient days at risk. Cox regression calculated the haz-
ard ratios and 95% confidence intervals for predictors of
first asthma readmission within 180 days. Time-to-event
TAGGEDEND56 HOGAN ET AL ACADEMIC PEDIATRICS
was defined as elapsed time from index hospitalization
discharge to asthma readmission or censoring at 180 days.
The proportional hazards assumption was tested. Back-
ward stepwise variable selection was used, with a 2-tailed
P<.05 considered to be significant. Clustering by hospi-
tal was assessed, but not observed and therefore, was not
adjusted for. The NRD provides sampling weights to cal-
culate national estimates; however, ANN typically does
not utilize sampling weights. As the goal was to compare
risk factors identified by traditional statistics and ANN,
sampling weights were not used. Logistic regression was
also used to directly compare traditional statistical meth-
ods with ANN, as both methods utilize binary outputs,
which can be compared using AUC. In addition, both
models had identical incomplete follow up for those
admitted in the second half of the year; given that the goal
was to compare risk factors between models this was
determined to be appropriate. Similar comparisons with
the Cox regression are difficult, due to the person-days at-
risk calculation. Both Cox and logistic regressions were
conducted using STATA 14.1 (College Station, Tex).
T
AGGEDH2ANN
We constructed an ANN to identify predictors of first
asthma readmission within 180 days. The ANN model was
constructed using a 50/50 split training and validation data-
set. We programmed a basic ANN, with 1 hidden layer
having 2 nodes, using the neuralnet function
16
found in the
R statistical software environment.
17
This uses back-propa-
gation to fit the model (the standard method of fitting such
a model), and multiple replications, considering all min-
max standardized independent variables and potential vari-
able interactions to predict readmissions. As the primary
outcome was binary (readmitted or censored), the sigmoid
function was used as the activation function in the hidden
and output layers. The back-propagation algorithm used a
learning rate of 0.05 and a momentum of 0.9. The learning
process was discontinued when the average error (mean
square error) in the training set decreased to 0.00001.
16
In light of the black-box nature of ANN, and to make the
outputs more useful for clinicians, we calculated the rela-
tive importance of the variables remaining in the ANN in
relation to their estimated weights in the model. In a regres-
sion model with standardized independent variables, the
coefficients of the fitted model can be interpreted as reflect-
ing the relative importance of the variables in question. In
ANN, standardization is applied twice: first, using the min-
max adjustment of the initial variables, and second, when
forcing fitted variables to lie in (0,1) by use of the sigmoi-
dal function. This standardization process suggests that a
similar interpretation to traditional regression coefficients
can be applied, though common outputs, such as odds
ratios, hazard ratios, or 95% confidence intervals, cannot
be derived. We calculated the weighted sum of the ANN
variable coefficients’ absolute values of the reported
weights, linking the input variables to the 2 “hidden layer”
(latent) variables. The absolute value of this weighted sum
was normalized and expressed on a percent basis. The
predictive accuracy of the ANN model was analyzed using
50/50 randomized training re-samples and cross-validation
techniques in 10 replications, and compared with the AUC
of the logistic regression model. There is currently no reli-
able ANN which can model time to event similar to Cox
regression, which is why ANN is only compared to logistic
regression.
TAGGEDH1RESULTSTAGGEDEND
The database yielded 18,489 pediatric asthma index
admissions and 1044 asthma readmissions, for a readmis-
sion rate of 5.7% at 180 days (Table 1). A total of 521 dis-
charges were excluded from analyses, due to missing data
(n = 341), transfer to another acute hospital (n = 81), leav-
ing against medical advice (n = 43), or death during hospi-
talization (n = 6). The median age at index admission was
8 years old (interquartile range, 6−12 years old); 41% of
index admissions were female patients, two thirds had
public insurance, 93% lived in a metropolitan area, and
6% had at least one CCC.
The Cox proportional-hazard model identified multiple
factors associated with time to readmission for pediatric
patients (Table 2a). Public insurance, prolonged LOS, and
nonwinter-month discharges were associated with higher
readmission risk, whereas micropolitan hospital location
was associated with lower risk. Logistic regression revealed
similar results, with public insurance, prolonged LOS, and
nonwinter-month discharges associated with higher read-
mission risk, and micropolitan hospital location associated
with lower risk (Table 2b). Overall model prediction was
weak,
18
with an AUC of 0.592 (Table 1a). There was no
meaningful AUC difference between the parsimonious
model and the traditional models including all variables.
In ANN, 9 factors and their complex interactions were
significantly associated with readmissions (Table 1b,
online). Four overlapped with the Cox model (LOS, prin-
cipal payer, season, and county size), whereas five were
unique (age, hospital bed number, teaching-hospital sta-
tus, weekend index admission, and CCCs). The normal-
ized importance of the variables is presented in Table 2.
Admission season was the most important factor, with a
standardized importance markedly higher than the next
most important risk factors. The ANN AUC was 0.637 in
the training set and 0.636 in the validation set, which was
greater than the AUC for logistic regression (Table 1b).
TAGGEDH1DISCUSSIONTAGGEDEND
In a nationally representative dataset, an ANN pre-
dicted 180-day readmissions marginally better than tradi-
tional statistical models. Although model prediction was
poor, both traditional and ANN models performed simi-
larly to models in other studies using retrospective data-
sets for asthma readmission prediction.
8
The Cox and logistic models showed that nonwinter
season, long LOS, public insurance, and micropolitan hos-
pitals were associated with 180-day asthma readmissions.
ANN also retained these variables in the model, and they
were the variables with the highest standardized
TAGGEDENDACADEMIC PEDIATRICS COMPARING ADVANCED ASTHMA READMISSION RISK FACTORS MODELING 57
importance. Admission season was an order of magnitude
more important than other risk factors in the ANN. In the
traditional models, although all nonwinter seasons were
associated with readmission, summer admission was asso-
ciated with a more that 2 and half times greater odds of
readmission versus winter admissions. This finding is con-
sistent with prior research, albeit not as strong an associa-
tion in previous studies.
19
Reflecting pediatric asthma
national seasonal trends, summer admissions were the
least frequent (13%), and winter admissions were most
frequent. It may be that the children prone to being admit-
ted in the summer months have worse chronic asthma
control, and therefore are more likely to be readmitted
when the fall and winter viruses begin circulating. This
association cannot be assessed without either prospective
data, more longitudinal data, or a more granular dataset
that includes asthma severity and markers of asthma con-
trol, such as the refill prescription ratio.
20
LOS of the
index hospitalization is also likely a marker of underlying
disease severity, and not surprisingly, was found to be
Table 1. 180-Day Readmission Rates by Patient and Hospital Characteristics
Characteristic Overall Cohort 180-Day Readmissions P
N, (%) 18,489 1044 (5.7) —
Age in years, % .106
5−11 74.4 4.5
12−18 25.6 6.1
Female gender, % 40.9 5.7 .765
Principal payer, % <.001
Private 28.5 4.8
Public 65.1 6.2
Uninsured 3.1 2.6
Other
†
3.3 6.2
Social risk factor, %
‡
0.3 0.05 .15
County population classification, % .135
Large metropolitan: ≥1 million 65.8 5.9
Small metropolitan: <1 million 27.3 5.4
Micropolitan 5.4 4.3
Non-urban 1.5 5.0
Weekend admission, % 27.7 5.6 .819
Index hospitalization LOS in days, % <.001
<2 33.6 4.5
2-3 51.7 6.0
≥4 14.7 7.1
Discharge season, % <.001
Winter (December−February) 24.2 3.5
Spring (March−May) 28.4 5.4
Summer (June−August) 13.2 7.8
Fall (September−November) 34.2 6.5
Hospital bed number, %
§
.303
Large 66.7 5.6
Medium 27.7 5.9
Small 5.6 4.8
Hospital teaching status, % .041
Urban nonteaching 19.6 5.2
Urban teaching 73.5 5.9
Nonurban 6.9 4.5
Hospital ownership, % .247
Private nonprofit 70.1 5.7
Government 20.1 5.8
Private, for-profit 9.8 4.8
Discharge disposition, % .352
Routine (home/home health care) 99.1 5.6
Skilled facilities 0.2 10.7
Other 0.2 5.7
Median household income*,% .172
Quartile 4: ≥$64,000 12.3 4.8
Quartile 3: $48,000−$63,999 18.6 5.7
Quartile 2: $38,000−$47,999 24.7 5.4
Quartile 1: $1−$37,999 44.4 6.0
Pediatric CCC, % 6.4 6.3 .293
LOS indicates length of stay; CCC, complex chronic condition.
*By patient's residential ZIP code.
†Other includes Worker's Compensation, CHAMPUS, CHAMPVA, Title V, and other government programs.
9
‡As identified by ICD-9 code.
12
§Size defined by hospital location and teaching status per NRD.
9
TAGGEDEND58 HOGAN ET AL ACADEMIC PEDIATRICS
associated with increased readmission in all of our models
and several prior studies.
21−23
Public insurance was significantly associated with read-
mission in both models. There are numerous factors which
are associated both with public insurance and pediatric
asthma readmissions but were not available in this dataset,
including residence in predominantly minority neighbor-
hoods,
24
residence in high-risk areas,
25
and race/
ethnicity,
4,26,27
so confounding is a distinct possibility. If
these variables were in the dataset, they likely would
improve both the predictive ability of a model using retro-
spective data, and allow deeper insight into the individual
risk factors found to associated with readmissions in the
models.
Hospital location in a micropolitan county was associ-
ated with a lower risk of readmission, compared with hos-
pitals in large cities. This finding is consistent with prior
research indicating lower pediatric asthma readmission
rates in rural counties versus metropolitan areas,
28
although one study suggested an increased 30-day read-
mission risk in micropolitan versus rural hospitals.
22
With
only 5% of patients admitted to micropolitan hospitals in
this dataset, this finding will need to be explored in future
research.
Table 2. Traditional Statistical Models for Factors Associated With 180-Day Pediatric Asthma Readmission in the 2013 Nationwide Read-
mission Database
2a: Cox Proportional-Hazards Model
Characteristic Hazard Ratio (95% CI) Adjusted Hazard Ratio (95% CI)
Principal payer
Private Referent Referent
Public 1.32 (1.14−1.52) 1.32 (1.14−1.52)
Uninsured 0.54 (0.32−0.90) 0.52 (0.31−0.88)
Other 1.32 (0.94−1.85) 1.33 (0.95−1.87)
County population classification
Large metropolitan: ≥1 million Referent Referent
Small metropolitan: <1 million 0.91 (0.79−1.04) 0.90 (0.78−1.52)
Micropolitan 0.72 (0.53−0.98) 0.69 (0.51−0.94)
Nonurban 0.84 (0.49−1.42) 0.84 (0.49−1.43)
Index hospitalization LOS, days
<2 Referent Referent
2−31.33 (1.15−1.53) 1.35 (1.17−1.56)
≥41.61 (1.34−1.93) 1.63 (1.36−1.96)
Discharge season
Winter (December−February) Referent Referent
Spring (March−May) 1.54 (1.26−1.87) 1.55 (1.28−1.89)
Summer (June−August) 2.33 (1.88−2.87) 2.40 (1.94−2.97)
Fall (September−November) 1.99 (1.66−2.40) 2.04 (1.69−2.45)
2b: Logistic Regression Model
Characteristic Odds Ratio (95% CI) Adjusted Odds Ratio (95% CI)
Principal payer
Private Referent Referent
Public 1.31 (1.13−1.52) 1.31 (1.13−1.52)
Uninsured 0.53 (0.32−0.91) 0.52 (0.31−0.89)
Other 1.33 (0.93−1.89) 1.34 (0.94−1.92)
County population classification
Large metropolitan: ≥1 million Referent Referent
Small metropolitan: <1 million 0.91 (0.79−1.05) 0.90 (0.78−1.04)
Micropolitan 0.72 (0.53−0.99) 0.70 (0.51−0.95)
Nonurban 0.84 (0.49−1.44) 0.83 (0.48−1.43)
Index hospitalization LOS, days
<2 Referent Referent
2−31.33 (1.15−1.55) 1.36 (1.17−1.57)
≥41.60 (1.33−1.94) 1.63 (1.35−1.97)
Discharge season
Winter (December−February) Referent Referent
Spring (March−May) 1.61 (1.32−1.97) 1.63 (1.34−1.99)
Summer (June−August) 2.38 (1.91−2.96) 2.45 (1.97−3.05)
Fall (September−November) 1.96 (1.62−2.36) 2.00 (1.65−2.41)
LOS indicates length of stay; CI, confidence interval.
Bold text indicates P<.05.
The parsimonious models above also considered all variables available in the NRD. Variables rejected included: age, gender, social risk-
factor score,
12
county size, weekend admission, hospital bed number, hospital teaching status, hospital ownership, discharge disposition,
median household income, and presence of a complex chronic condition.
TAGGEDENDACADEMIC PEDIATRICS COMPARING ADVANCED ASTHMA READMISSION RISK FACTORS MODELING 59
There was consistent concordance of all three models
on 4 readmission risk factors: index admission in a non-
winter season, LOS, public insurance, and micropolitan
hospitals. These findings have important implications for
practice, research, and policy. For example, the study
results suggest that Medicaid-covered children being dis-
charged in the summer and after a prolonged LOS would
benefit from intensive case management focused on
reducing readmission risk. Health systems and policy
makers should invest in evidenced-based asthma readmis-
sion reduction solutions, such as parent mentors,
29
or mul-
ticomponent interventions,
7
and implement these
programs in the highest risk groups for efficiency.
Readmission risk factors uniquely identified by ANN
deserve further consideration. CCCs are well known to
have an increased risk of morbidity, including prior
research evaluating multiple asthma readmission time
frames
10
that includes 180 days.
30,31
Similarly, older age
and hospital teaching status were associated with readmis-
sion in the ANN and prior studies.
32
To our knowledge,
hospital bed number and weekend admission have not
been previously identified as increased readmission risk
factors in pediatric asthma. Excepting bed number and
weekend admission, these risk factors have been regularly
and robustly found to be associated with pediatric asthma
readmissions. This suggests that traditional modeling tech-
niques may not be able to identify higher-level connections
between multiple variables or subtler variable interactions
which ANN can capture. Differences in identified risk fac-
tors between standard statistical methods and ANN may
be due to the ability of the ANN to account for the com-
plex connections among the independent variables, which
is not possible in traditional statistical models. Alternate
machine-learning-based techniques, including recursive
partitioning analyses, random forest, boosted gradient, and
support vectors, may also be superior, depending on the
dataset, for producing maximal sensitivity, specificity, pre-
dictive value, and ease of use for clinicians.
Certain study limitations should be noted. NRD lacks
longitudinal outcomes for those admitted in the second
half of the year, as NRD does not maintain case identifiers
across annual datasets, due to inherent sampling strategies
and privacy protections. This limitation is partially miti-
gated by accounting for the admission season and using
patient days at-risk with Cox proportional-hazards model-
ing. The 2013 NRD uses ICD-9 coding, which has been
replaced by ICD-10; future research is planned to use
these newer diagnosis codes. Another limitation is the
lack of certain risk factors in the dataset known to be
associated with increased asthma readmissions. Prior
health care utilization, such as prior-year (before the index
admission) hospitalizations, ED visits, intensive-care-unit
admissions, and outpatient follow-up, are not reported in
NRD, but are associated with increased pediatric asthma
readmission risk.
33
Prescription claims data, such as the
asthma medication ratio (number of controller medication
claims/[number of controller medication claims + number
of rescue medication claims]) would likely also augment
readmission predictions.
34
NRD reports the number of
hospital beds, but does not differentiate children’s hospi-
tals from community hospitals, which have different read-
mission rates.
35
NRD does not report race/ethnicity,
which is problematic, as African-American children have
increased rates of asthma readmissions, and Latino child-
rens’ asthma prevalence and exacerbation risks are high
and vary by subgroups.
36
Similarly, measures of racism,
economic stability, housing issues, English proficiency,
immigration status, and other social determinants of health
are not present in the NRD, and are rarely included in
administrative datasets, which may partially explain the
poor predictive performance of models based on retro-
spective databases.
8,37,38
Routine incorporation of these
factors into administrative and clinical datasets will allow
for more nuanced investigations of readmission risk fac-
tors and likely lead to more robust predictive models.
Late pediatric asthma readmissions are caused by a web
of factors: underlying disease severity, the transition from
hospital to home, neighborhood factors, and many
others.
39
The variables in the NRD, and administrative
datasets generally, are pale reflections of the underlying
readmission causes. Although ANN did outperform tradi-
tional models, the accuracy of the prediction models was
hamstrung by the lack of richer, more granular data. The
overlapping risk factors between all three models are
likely the most robust readmission predictors, at least for
now, and should be considered in any future readmission
reduction efforts.
TAGGEDH1CONCLUSIONSTAGGEDEND
Different methods can produce different readmission
models. Relying on traditional modeling alone may over-
look key readmission risk factors and complex factor
interactions identified by neural networks. Four risk fac-
tors—nonwinter-month admission, long LOS, public
insurance, and micropolitan hospitals—were shared by all
3 models, and could prove particularly powerful for pre-
dicting readmissions. Five risk factors were uniquely
identified by ANN: age, hospital bed number, teaching-
hospital status, weekend index admission, and CCCs.
Greater insight is needed on how to harmonize dissonant
findings of different models so that children at greatest
readmission risk can be identified.
TAGGEDH1ACKNOWLEDGMENTSTAGGEDEND
Financial statement: Dr Hogan was supported by the Connecticut
Institute for Clinical and Translational Science at the University of Con-
necticut, and an Academic Pediatric Association’s Young Investigator
Award. The study sponsors had no role in study design, the collection,
analysis, interpretation of data, the writing of the report, or the decision
to submit the paper for publication. The content is solely the responsibil-
ity of the authors, and does not necessarily represent the official views of
either awarding institution.
TAGGEDH1REFERENCESTAGGEDEND
1. Kamble S, Bharmal M. Incremental direct expenditure of treating
asthma in the United States. J Asthma. 2009;46:73–80. https://doi.
org/10.1080/02770900802503107.
T
AGGEDEND60 HOGAN ET AL ACADEMIC PEDIATRICS
2. Chang LV, Shah AN, Hoefgen ER, et al. Lost earnings and nonmed-
ical expenses of pediatric hospitalizations. Pediatrics. 2018;142:
e20180195. https://doi.org/10.1542/peds.2018-0195.
3. Hain PD, Gay JC, Berutti TW, et al. Preventability of early readmis-
sions at a children’s hospital. Pediatrics. 2013;131:e171–e181.
https://doi.org/10.1542/peds.2012-0820.
4. Parikh K, Berry J, Hall M, et al. Racial and ethnic differences in
pediatric readmissions for common chronic conditions. J Pediatr.
2017;186:158–164. https://doi.org/10.1016/j.jpeds.2017.03.046.
Published online.
5. Bergert L, Patel SJ, Kimata C, et al. Linking patient-centered
medical home and asthma measures reduces hospital readmission
rates. Pediatrics. 2014;134:e249–e256. https://doi.org/10.1542/
peds.2013-1406.
6. Lu S, Kuo DZ. Hospital charges of potentially preventable pediatric
hospitalizations. Acad Pediatr. 2012;12:436–444. https://doi.org/
10.1016/j.acap.2012.06.006.
7. Parikh K, Keller S, Ralston S. Inpatient quality improvement inter-
ventions for asthma: a meta-analysis. Pediatrics. 2018;141:
e20173334. https://doi.org/10.1542/peds.2017-3334.
8. Kansagara D, Englander H, Salanitro A, et al. Risk prediction mod-
els for hospital readmission: a systematic review. JAMA.
2011;306:1688–1698. https://doi.org/10.1001/jama.2011.1515.
9. Healthcare Cost and Utilization Project (HCUP). HCUP NRD
description of data elements. Agency for Healthcare Research and
Quality. Available at: www.hcup-us.ahrq.gov/db/vars/zipinc_qrtl/
nrdnote.jsp. Accessed January 5, 2020.
10. Veeranki SP, Ohabughiro MU, Moran J, et al. National estimates of
30-day readmissions among children hospitalized for asthma in the
United States. J Asthma. 2018;55:695–704. https://doi.org/10.1080/
02770903.2017.1365888.
11. Morosco G, Kiley J, National Heart Lung and Blood Institute. Expert
panel report 3 (EPR-3): guidelines for the diagnosis and management
of asthma-summary report 2007. J Allergy Clin Immunol. 2007;120
(suppl 5):S94–S138. https://doi.org/10.1016/j.jaci.2007.09.043.
12. Pantell MS, Kaiser SV, Torres JM, et al. Associations between
social factor documentation and hospital length of stay and readmis-
sion among children. Hosp Pediatr. 2020;10:12–19. https://doi.org/
10.1542/hpeds.2019-0123.
13. United States Department of Agriculture. Urban influence codes.
USDA Data Products. 1993. Available at: https://www.ers.usda.
gov/data-products/urban-influence-codes.aspx Accessed January 5,
2020.
14. Feudtner C, Feinstein JA, Zhong W, et al. Pediatric complex chronic
conditions classification system version 2: updated for ICD-10 and
complex medical technology dependence and transplantation. BMC
Pediatr. 2014;14:199. https://doi.org/10.1186/1471-2431-14-199.
15. Healthcare Cost and Utilization Project. NRD Overview. 2017.
Available at: www.hcup-us.ahrq.gov. Accessed August 27, 2021.
16. Fritsch S, Guenther F, Wright MN, et al. Neuralnet: Training of
Neural Networks. 20192019:1–15. Available at: https://cran.r-proj
ect.org/web/packages/neuralnet/neuralnet.pdf. Accessed August 27,
2021.
17. R Core Team. R: A Language and Enviorment for Stasticical Comput-
ing. Vienna, Austria: R Foundation for Statistical Computing; 2020.
Available at: https://www.R-project.org/. Accessed August 27, 2021.
18. Hosmer Jr DW, Lemeshow S, Sturdivant RX. Applied Logistic
Regression. 398. John Wiley & Sons; 2013. https://doi.org/10.4135/
9781412983433.
19. Brittan M, Richardson T, Kenyon C, et al. Association between
postdischarge oral corticosteroid prescription fills and readmission
in children with asthma. J Pediatr. 2017;180:163–169. https://doi.
org/10.1016/j.jpeds.2016.09.034.
20. Broder MS, Gutierrez B, Chang E, et al. Ratio of controller to total
asthma medications: determinants of the measure. Am J Manag
Care. 2010;16:170–178.
21. Knighton AJ, Flood A, Speedie SM, et al. Does initial length of
stay impact 30-day readmission risk in pediatric asthma
patients? J Asthma. 2013;50:821–827. https://doi.org/10.3109/
02770903.2013.816726.
22. Veeranki SP, Ohabughiro MU, Moran J, et al. National estimates of
30-day readmissions among children hospitalized for asthma in the
United States. J Asthma. 2018;55:695–704. https://doi.org/10.1080/
02770903.2017.1365888.
23. Hogan AH, Carroll CL, Iverson MG, et al. Risk factors for pediatric
asthma readmissions: a systematic review. J Pediatr.
2021;236:219–228. https://doi.org/10.1016/j.jpeds.2021.05.015.
24. Liu SY, Pearlman DN. Hospital readmissions for childhood asthma:
the role of individual and neighborhood factors. Public Health Rep.
2009;124:65–78.
25. Beck AF, Simmons JM, Huang B, et al. Geomedicine: area-based
socioeconomic measures for assessing risk of hospital reutilization
among children admitted for asthma. Am J Public Health.
2012;102:2308–2314. https://doi.org/10.2105/AJPH.2012.300806.
26. Beck AF, Huang B, Auger KA, et al. Explaining racial disparities in
child asthma readmission using a causal inference approach. JAMA
Pediatr. 2016;170:695–703. https://doi.org/10.1001/jamapediat-
rics.2016.0269.
27. Kenyon CC, Melvin PR, Chiang VW, et al. Rehospitalizat ion fo r
childhood asthma: timing, variation, and opportunities for inter-
vention. JPediatr. 2014;164:300–305. https://doi.org/10.1016/j.
jpeds.2013.10.003.
28. Rushworth RL, Rob MI. Readmissions to hospital: the contribution
of morbidity data to the evaluation of asthma management. Aust J
Public Health. 1995;19:363–367.
29. Flores G, Bridon C, Torres S, et al. Improving asthma outcom es in
minority children: a randomized, controlled trial of parent men-
tors. Pediatrics. 2009;124:1522–1532. https://doi.org/10.1542/
peds.2009-0230.
30. Kenyon CC, Rubin DM, Zorc JJ, et al. Childhood asthma hos-
pital discharge medication fills and risk of subsequent readmis-
sion. J Pediatr. 2015;166:1121–1127. https://doi.org/10.1016/j.
jpeds.2014.12.019.
31. Kenyon CC, Melvin PR, Chiang VW, et al. Rehospitalization for
childhood asthma: timing, variation, and opportunities for interven-
tion. J Pediatr. 2014;164:300–305. https://doi.org/10.1016/j.
jpeds.2013.10.003.
32. Chen Y, Dales R, Stewart P, et al. Hospital readmissions for asthma
in children and young adults in Canada. Pediatr Pulmonol.
2003;36:22–26. https://doi.org/10.1002/ppul.10307.
33. Auger KA, Kahn RS, Davis MM, et al. Pediatric asthma readmis-
sion: asthma knowledge is not enough? J Pediatr. 2015;166:101–
108. https://doi.org/10.1016/j.jpeds.2014.07.046.
34. Andrews AL, Brinton DL, Simpson KN, et al. A longitudinal
examination of the asthma medication ratio in children with
Medicaid. JAsthma. 2020;57:1083–1091. https://doi.org/
10.1080/02770903.2019.1640727.
35. Silber JH, Rosenbaum PR, Even-shoshan O, et al. Length of stay,
conditional length of stay, and prolonged stay in pediatric asthma.
Health Serv Res. 2003;38:867–886. https://doi.org/10.1111/1475-
6773.00150.
36. Lara M, Akinbami L, Flores G, et al. Heterogeneity of childhood
asthma among Hispanic children: Puerto Rican children bear a dis-
proportionate burden. Pediatrics. 2006;117:43–53. https://doi.org/
10.1542/peds.2004-1714.
37. Hogan AH, Flores G. Social determinants of health and the hospital-
ized child. Hosp Pediatr. 2020;10:101–103. https://doi.org/10.1542/
hpeds.2019-0289.
38. Flores G, Hollenbach JP, Hogan AH. To eliminate racial and ethnic
disparities in child health care, more needs to be addressed than just
social determinants. Lancet Respir Med. 2019;7:842–843. https://
doi.org/10.1016/S2213-2600(19)30306-6.
39. Nakamura MM, Toomey SL, Zaslavsky AM, et al. Measuring pedi-
atric hospital readmission rates to drive quality improvement. Acad
Pediatr. 2014;14(suppl 5):S39–S46. https://doi.org/10.1016/j.
acap.2014.06.012.
TAGGEDENDACADEMIC PEDIATRICS COMPARING ADVANCED ASTHMA READMISSION RISK FACTORS MODELING 61