Who’s your data? Primary immune deficiency differential diagnosis prediction via machine learning and data mining of the USIDNET registry

Abstract

INTRODUCTION: There are currently more than 450 primary immune deficiency (PID) diseases, and about 7,000 rare diseases that together afflict around 1 in every 17 humans. Computational aids based on data mining and machine learning might facilitate the diagnostic task by extracting rules from large datasets and making predictions when faced with new problem cases.
OBJECTIVE: In a proof-of-concept data mining study, we aimed to predict PID diagnoses with a supervised machine learning algorithm based on classification tree boosting.
METHODS: Through a data query at the USIDNET registry we obtained a database of 2,396 patients with common diagnoses of PID, including their clinical and laboratory features. We kept 12 diagnoses and 286 features that were included in the model. We used the XGBoost package with parallel tree boosting for the supervised classification model, and SHAP for variable importance interpretation, on Python v3.7. The patient database was split into training and testing subsets, and after boosting through gradient descent, the predictive model provides measures of diagnostic prediction accuracy and individual feature importance. To correct for imbalanced classification, after a baseline performance test, we used the Class Weighting Hyperparameter, or scale_pos_weight.
RESULTS: The twelve PID diagnoses were CVID (1,098 patients), DiGeorge syndrome, Chronic granulomatous disease, Congenital agammaglobulinemia, ID not otherwise classified, Specific antibody deficiency, Complement deficiency, Hyper-IgM, Leukocyte adhesion deficiency, ectodermal dysplasia with immune deficiency, Severe combined immune deficiency, and Wiskott-Aldrich syndrome. For CVID, the model found an accuracy on the train sample of 0.80, with an area under the ROC curve (AUC) of 0.80, and a Gini coefficient of 0.60. In the test subset, accuracy was 0.76, AUC 0.75, and Gini 0.51. The positive feature value to predict CVID was highest for upper respiratory infections, asthma, autoimmunity and hypogammaglobulinemia. Features with the highest negative predictive value were high IgE, growth delay, abscess, lymphopenia, and congenital heart disease. For the rest of the diagnoses, accuracy stayed between 0.75 and 0.99, AUC 0.46-0.87, Gini 0.07-0.75, and LogLoss 0.09-8.55. See tables and figures.
DISCUSSION: Clinicians should remember to consider the negative predictive features together with the positives. We are calling this a proof-of-concept study to continue with our explorations. A good performance is encouraging, and the feature importance might aid feature selection for future endeavors. In the meantime, we can learn from the rules derived by the model and build a user-friendly decision tree to generate differential diagnoses.

Full text

Who’s your data?
Primary immune deficiency differential diagnosis prediction
via machine learning and data mining of the USIDNET registry
Jose Alfredo Méndez Barrera1, Samuel Rocha Guzmán1, Elisa Hierro Cascajares2, Elizabeth
K Garabedian3, Ramsay L Fuleihan4, Kathleen E Sullivan5, and Saul O Lugo Reyes2
From the (1) Data Science Department, Autonomous Technological Institute of Mexico,
Mexico City, Mexico; (3) National Institutes of Health, Bethesda, MD, USA; (4) Division
of Pediatric Allergy, Immunology and Rheumatology at Columbia University, New York
City, NY, USA; (5) Children’s Hospital of Philadelphia, PA, USA; and (2) Immune
deficiencies Lab, National Institute of Pediatrics, Secretariat of Health, Mexico City,
Mexico
Address correspondence to: dr.lugo.reyes@gmail.com
KEYWORDS: Rare diseases, inborn errors of immunity, data mining, extreme gradient
boosting, machine learning, diagnosis prediction, classification, primary immune
deficiencies, registry.

ABSTRACT:
INTRODUCTION: There are currently more than 450 primary immune deficiency (PID)
diseases, and about 7,000 rare diseases that together afflict around 1 in every 17 humans.
Computational aids based on data mining and machine learning might facilitate the
diagnostic task by extracting rules from large datasets and making predictions when faced
with new problem cases.
OBJECTIVE: In a proof-of-concept data mining study, we aimed to predict PID diagnoses
with a supervised machine learning algorithm based on classification tree boosting.
METHODS: Through a data query at the USIDNET registry we obtained a database of
2,396 patients with common diagnoses of PID, including their clinical and laboratory
features. We kept 12 diagnoses and 286 features that were included in the model. We used
the XGBoost package with parallel tree boosting for the supervised classification model,
and SHAP for variable importance interpretation, on Python v3.7. The patient database was
split into training and testing subsets, and after boosting through gradient descent, the
predictive model provides measures of diagnostic prediction accuracy and individual
feature importance. To correct for imbalanced classification, after a baseline performance
test, we used the Class Weighting Hyperparameter, or scale_pos_weight.
RESULTS: The twelve PID diagnoses were CVID (1,098 patients), DiGeorge syndrome,
Chronic granulomatous disease, Congenital agammaglobulinemia, ID not otherwise
classified, Specific antibody deficiency, Complement deficiency, Hyper-IgM, Leukocyte
adhesion deficiency, ectodermal dysplasia with immune deficiency, Severe combined
immune deficiency, and Wiskott-Aldrich syndrome. For CVID, the model found an
accuracy on the train sample of 0.80, with an area under the ROC curve (AUC) of 0.80, and
a Gini coefficient of 0.60. In the test subset, accuracy was 0.76, AUC 0.75, and Gini 0.51.
The positive feature value to predict CVID was highest for upper respiratory infections,
asthma, autoimmunity and hypogammaglobulinemia. Features with the highest negative
predictive value were high IgE, growth delay, abscess, lymphopenia, and congenital heart
disease. For the rest of the diagnoses, accuracy stayed between 0.75 and 0.99, AUC 0.46-
0.87, Gini 0.07-0.75, and LogLoss 0.09-8.55. See tables and figures.

DISCUSSION: Clinicians should remember to consider the negative predictive features
together with the positives. We are calling this a proof-of-concept study to continue with
our explorations. A good performance is encouraging, and the feature importance might aid
feature selection for future endeavors. In the meantime, we can learn from the rules derived
by the model and build a user-friendly decision tree to generate differential diagnoses.
ABBREVIATIONS: AI, artificial intelligence; AUC, Area under the ROC curve; CGD,
chronic granulomatous disease; CVID, common-variable immune deficiency; DGS,
DiGeorge syndrome; PID, primary immune deficiencies; ROC, Receiver Operating
Characteristic; SCID, severe combined immune deficiency; SHAP, Shapley additive
explanations; USIDNET, United States Immunodeficiency Network; XGBoost, extreme
gradient boosting.

INTRODUCTION:
Inborn errors of immunity, also known as Primary immune deficiencies (PID), are a
heterogeneous group of over 450 congenital rare diseases with increased susceptibility to
infection, inflammation, autoimmunity, allergy and/or cancer (1). Their estimated global
prevalence is 1 in 10,000 humans. The United States Immunodeficiency Network
(USIDNET) is a research consortium that maintains a registry of data from over 4,000
patients with PID (usidnet.org).
Patients with rare diseases usually endure an odyssey of several years before
reaching a correct diagnosis (2). Diagnostic errors are an important cause of deaths,
complaints, complications, and waste (3). The human gene connectome at Rockefeller
University has predicted a ceiling of more than 3,100 genes involved in the immune system
(4). In the following decade, gene discovery by next-generation sequencing could bring the
number of PID diseases to the thousands, one-upping the complexity of the diagnostic task
as well.
Computers might help clinicians diagnose and reduce diagnostic errors (5),
especially in the field of primary immune deficiencies and other rare diseases, which are
usually congenital, with an early and syndromic presentation (6). Artificial intelligence
(A.I.) solutions are already implemented in financial services, global
positioning/navigation, and biological sciences, among other human endeavors.
Data mining and machine learning are two branches of A.I. that make use of
existing large datasets to find patterns and extract rules with algorithms or automated

methods that make sense of data to then make predictions about new data sets or problem
cases.
Extreme Gradient Boosting (XGBoost) is a robust supervised machine learning
algorithm (https://xgboost.readthedocs.io/en/latest/). Classification tree boosting consists of
the generation of multiple sequential “weak” prediction models, each of which is taken to
generate a stronger, more stable model by tweaking the results of the previous model. This
optimization algorithm is called a Gradient descent. The average gain of each feature in all
trees in which it appears is used to rank the features in importance (7). Hyperparameter
optimization, or tuning, is the measure and choosing of the most valuable parameters, by
giving them different weights or constraints, thus improving learning rates and minimizing
function losses to optimize the model.
For the interpretation of the importance of variables, SHAP (Shapley Additive
exPlanations) can be helpful (https://github.com/slundberg/shap). This method assigns to
each feature an importance value for a unique prediction and facilitates the global and local
interpretability of any machine learning model, thus making for a more user-friendly
interpretation (8).
OBJECTIVE:
In this proof-of-concept data mining study, we aimed to predict PID diagnoses using
a machine learning algorithm.

METHODS:
Through a data query at the USIDNET registry (usidnet.org), we obtained a
database of 2,396 patients with common diagnoses of primary immune deficiencies (PID),
including their clinical and laboratory features. The database was curated and re-codified to
reduce dimensionality and convert continuous variants into categorical (e.g.,
high/low/normal IgG); we kept 12 diagnoses (rows) and 286 features (columns) that were
included in the model.
We used the Extreme Gradient Boost (XGBoost) package with parallel tree boosting
for the supervised classification model, and SHAP (Shapley Additive exPlanations) for
variable importance interpretation, on Python version 3.7 (Jupyter) and Google
Colaboratory notebooks.
Each diagnosis is differentiated or predicted against all others (i.e., CVID versus 11
other diagnoses). The patient database is split into training and testing subsets, and after
boosting through gradient descent, the predictive model provides measures of prediction
probability: accuracy, area under the ROC curve (AUC), gini index, and Logloss. To
correct for imbalanced classification, after a baseline performance test, we used the Class
Weighting Hyperparameter, or scale_pos_weight for all diagnoses except the most
abundant: CVID and DiGeorge Syndrome.
This study design is exempt from Institutional Review Board approval at the
National Institute of Pediatrics. We maintain and respect the confidentiality of patients and

their families. All authors subscribe and uphold the Helsinki Declaration involving
experimentation on human subjects.
RESULTS:
The twelve PID diagnoses were CVID (1,098 patients), DiGeorge syndrome (DGS,
406), Chronic granulomatous disease (154), Congenital agammaglobulinemia (135), Not
otherwise classified (CORE, 132), Specific antibody deficiency (SAD, 117), Complement
deficiency (12), Hyper-IgM (HIGM, 46), Leukocyte adhesion deficiency (6), Ectodermal
dysplasia with immune deficiency (NEMO, 25), Severe combined immune deficiency
(SCID, 202), and Wiskott-Aldrich syndrome (WAS, 63 cases). See figure 1.
For the most frequent diagnosis, Common variable immune deficiency (CVID), the
model found an accuracy on the train sample of 0.80, with an area under the ROC curve
(AUC) of 0.80, and a Gini coefficient of 0.60. In the test subset, accuracy was 0.76, AUC
0.75, and Gini 0.51. The positive feature value to predict CVID was highest for upper
respiratory infections, asthma, autoimmunity and hypogammaglobulinemia. Features with
the highest negative predictive value were high IgE, growth delay, abscess, lymphopenia,
and congenital heart disease. See figure 2.
For DiGeorge Syndrome (DGS, n=406), the accuracy on the test sample was 0.90,
with an AUC of 0.87, and gini index of 0.75. Feature importance was highest for congenital
heart disease, facial dysmorphism, palate malformation, and hydroelectrolytic disorders.
Top negative features include Low IgA, upper respiratory infections, abscess, diarrhea, and
low IgG. Figure 3.

Specific antibody deficiency (SPAD, n=117) had an accuracy on test of 0.84, AUC
0.74, and gini of 0.49. Important positive features were Upper respiratory and ear
infections, allergic rhinitis, and Streptococcus; top negative features: Low IgG, Low IgM,
cardiovascular disease, teeth, diarrhea, eczema, gastrointestinal manifestations, and High
IgE.
For Agammaglobulinemia (AGAMMA, n=135), accuracy was 0.89, AUC 0.76, gini
0.519; Logloss was 3.84. Top positive features were Low IgM and IgG, Ear, Eye, and Bone
infections, and arthritis; top negatives were Candida sp., cardiovascular disease,
leukopenia, endocrine-metabolic disorders, and lymph nodes.
For Chronic granulomatous disease (CGD, n=154), accuracy was 0.92, AUC 0.83,
gini 0.67, and Logloss 2.69. Top positive predictive features were Abscess, high IgA,
Aspergillus sp, lymph nodes, and inflammatory bowel disease. Top negative predictive
features: Upper respiratory infections, Low IgM, Asthma, Low IgM, and cytopenias.
After target-variable (y) scaling to correct for imbalanced classification,
Complement deficiencies (COMPDEF, n=12) had a prediction accuracy of 0.92, AUC of
0.46, gini of -0.07, and a Logloss of 2.64. Positive predictive features were Skin infection,
Upper respiratory, and Mouth infection. Negative: Growth delay, cardiovascular disease,
low IgA, Cytopenias, endocrine-metabolic disorders, bronchiectasis, and abscess. See
tables and figures for accuracy, numbers, AUC, Gini and LogLoss for the rest of diagnoses.

DISCUSSION:
We found a good performance of XGBoost to predict any of twelve PID diagnosis
in a dataset of over 2,300 patients from the USIDNET registry, with accuracy and AUC
between 0.70 and 0.80 for most diseases, and Gini indexes above 0.50.
The strengths of our approach include a large total number of diagnosed patients, a
robust machine learning algorithm, and a random train/test split step for cross-validation.
The main limitations are a small list of diagnoses, a lack of non-PID diagnoses, and the lack
of a prospective validation study.
In the past few years, we have toyed with different approaches to machine learning
classification, including linear discriminant analysis, decision tree and random forests (9).
In the field of PID diseases, one group in Finland (10–12) and another in Houston at Baylor
(13) have attempted machine learning-assisted classification and prediction models, with
promising results.
In 2020 at Wuhan, Gao and Ding compared machine learning strategies to predict
breast cancer (BC) and cardiovascular disease (CVD) in datasets from the Irving repository
(14). They found a 94.74% accuracy for the XGBoost model in BC patients, with the fine
needle image of breast lump as the most important predictive value, and a 73.5% accuracy
for the CVD dataset XGBoost model, with systolic blood pressure as the most important
feature.
There are not many implications for clinical practice yet. The astute clinician should
remember to consider the negative predictive features, or counterfactuals, together with the

positives. We are calling this a proof-of-concept study to continue with our explorations. A
good performance is encouraging, and the feature importance might aid feature selection
for future endeavors.
Next, we want to develop a supervised decision tree or expert system that will make
use of both our knowledge base and the most important features from this study, in a user-
friendly questionnaire with consecutive classifying questions to predict the genetic
diagnosis of new problem cases with suspected PID.
A computational aid to assist the clinical diagnosis of rare diseases might facilitate
the work of physicians. Encouraging preliminary results from this and other studies suggest
such an aid is feasible to develop. Today still, A.I. stands for both Artificial Intelligence
and Almost Implemented.
ACKNOWLEDGEMENTS:
The U.S. Immunodeficiency Network (USIDNET), a program of the Immune
Deficiency Foundation (IDF), is supported by a cooperative agreement, U24AI86837, from
the National Institute of Allergy and Infectious Diseases (NIAID). The USIDNET
Consortium is composed of over a hundred clinicians who have contributed individually
with one to hundreds of patient registrations and their features, available at the registry.

DECLARATIONS
Funding: No funding was received for this manuscript.
CONFLICT OF INTERESTS: All authors declare no competing interests to disclose.
Availability of data and material: All the data supporting this study is available upon
request.
Code availability: Not applicable. The algorithms, language, workflow, and platforms used
are open-source software. Should the reader be interested in a verbatim sequence of steps,
we can provide it upon request.
Consent to participate: Not applicable, the study does not involve human subjects.
Authors contributions: JAMB converted values, depurated the database and reduced
dimensionality, applied the algorithm with SRG, analyzed results, critically appraised the
rough draft, and approved the final version of the manuscript. SRG coordinated the
application of the algorithm, proposed a solution for imbalanced classification, analyzed
results, critically appraised the rough draft, and approved the final version of the
manuscript. EHC curated the database, helped depurate and reduce features, built the
multidisciplinary team, and approved the final version of the manuscript. EKG contributed
with the capture of data for over ten percent of patients in the database, critically appraised
the rough draft, and approved the final version of the manuscript. RLF also contributed
with the capture of data for over ten percent of patients in the database and approved the
final version of the manuscript. KES contributed with over ten percent of patients in the
database, established contact with USIDNET, facilitated the international collaboration,
critically appraised the rough draft, and approved the final version of the manuscript. The
USIDNET Consortium contributed data from thousands of patients with PID, available at
the registry. SOLR conceived the study and the publication, wrote the manuscript and
coordinated the joint effort.
Consent for publication: from all authors and the USIDNET Consortium.
IRB Ethics Approval: Exempt study. We protect the confidentiality of patients and
families.

REFERENCES:
1. Tangye SG, Al-Herz W, Bousfiha A, Chatila T, Cunningham-Rundles C, Etzioni A, et
al. Human Inborn Errors of Immunity: 2019 Update on the Classification from the
International Union of Immunological Societies Expert Committee. J Clin Immunol.
2020 Jan 1;40(1):24–64.
2. Tan TY, Dillon OJ, Stark Z, Schofield D, Alam K, Shrestha R, et al. Diagnostic
impact and cost-effectiveness of whole-exome sequencing for ambulant children
with suspected monogenic conditions. JAMA Pediatr. 2017 Sep 1;171(9):855–62.
3. Makary MA, Daniel M. Medical error-the third leading cause of death in the US.
BMJ. 2016 May 3;353:i2139.
4. Itan Y, Casanova J-L. Novel Primary Immunodeficiency Candidate Genes Predicted
by the Human Gene Connectome. Front Immunol. 2015;6(April):1–8.
5. Segal M. How doctors think, and how software can help avoid cognitive errors in
diagnosis. Acta Paediatr [Internet]. 2007/09/14. 2007;96(12):1720–2. Available
from: http://www.ncbi.nlm.nih.gov/pubmed/17850393
6. Berman JJ. Rare diseases and orphan drugs. First Edit. San Diego, CA, USA:
Elsevier; 2014. 354 p.
7. Tutorial: XGBoost en Python. XGBoost (Extreme Gradient Boosting), es… | by Juan
Bosco Mendoza Vega | Medium [Internet]. [cited 2021 Oct 29]. Available from:
https://medium.com/@jboscomendoza/tutorial-xgboost-en-python-53e48fc58f73
8. SHAP for explainable machine learning [Internet]. [cited 2021 Oct 29]. Available
from: https://meichenlu.com/2018-11-10-SHAP-explainable-machine-learning/
9. Murata C, Ramirez A, Ramirez G, Cruz A, Morales J, Lugo-Reyes S. Análisis
discriminante para predecir el diagnóstico clínico de inmunodeficiencias primarias:
reporte preliminar. Rev Alerg México. 2015;(62):125–33.
10. Samarghitean C, Iltanen K, Juhola M, Vihinen M, Rugg G. Machine learning
methods for primary immunodeficiency diagnosis. In: 17th biennial meeting of the
European Society for Immunodeficiencies, Barcelona. 2016.
11. Samarghitean C. PIDexpert-decision support system for primary
immunodeficiencies. [online] [Internet]. University of Tampere. 2008. Available from:
http://bioinf.uta.fi/PIDexpert
12. Samarghitean C, Ortutay C, Vihinen M. Systematic classification of primary
immunodeficiencies based on clinical, pathological, and laboratory parameters. J
Immunol [Internet]. 2009 Dec 1 [cited 2013 Nov 8];183(11):7569–75. Available from:
http://www.ncbi.nlm.nih.gov/pubmed/19917694
13. Rider NL, Cahill G, Motazedi T, Wei L, Kurian A, Noroski LM, et al. PI prob: A risk
prediction and clinical guidance system for evaluating patients with recurrent
infections. PLoS One [Internet]. 2021;16(2 February):1–15. Available from:
http://dx.doi.org/10.1371/journal.pone.0237285

14. Gao L, Ding Y. Disease prediction via Bayesian hyperparameter optimization and
ensemble learning. BMC Res Notes [Internet]. 2020 Apr 10 [cited 2021 Oct
29];13(1). Available from: /pmc/articles/PMC7146897/

TABLES AND FIGURES
Figure 1. List of 12 PIDD diagnoses and number of patients from the USIDNET registry.
Each target diagnosis (1) is predicted against all others (0).
AB deficiency: Specific polysaccharide antibody deficiency; AGAMMA: Congenital
agammaglobulinemia; CGD: Chronic granulomatous disease, COMPDEF: complement
deficiencies; CORE: immune deficiency, not otherwise specified; CVID: common variable
immune deficiency; DGS, DiGeorge syndrome; HIGM, Hyper-IgM syndrome; LAD,
Leukocyte adhesion deficiency; NEMO, X-linked ectodermal dysplasia with
immunodeficiency; SCID, Severe-combined immune deficiency; WAS, Wiskott-Aldrich
syndrome. Table from Google Colab.

Table 1. Performance measures for the 12 diagnoses predictions, after random train/test
split. In all but CVID and DGS, the Class Weight Hyperparameter was used to correct for
imbalanced classification.
PIDD n Accuracy AUC Gini LogLoss
CVID 1098 0.75 0.75 0.49 8.55
DGS 406 0.897 0.87 0.75 3.55
SCID 202 0.85 0.73 0.47 5.14
CGD 154 0.92 0.83 0.66 2.68
AGAMMA 135 0.88 0.76 0.52 3.84
CORE (ID) 132 0.88 0.72 0.44 4.18
SPAD 117 0.84 0.74 0.49 5.57
WAS 63 0.76 0.74 0.48 8.31
HIGM 46 0.93 0.65 0.30 2.59
NEMO-ID 25 0.97 0.62 0.23 1.01
COMPDEF 12 0.92 0.46 -0.07 2.64
LAD 6 0.997 0.66 0.33 0.09

Figure 2. Twenty most important features to predict the diagnosis CVID and their SHAP
values. Higher and red signify greater importance; right/left indicate presence or absence,
respectively.
Alto: high, Bajo: low.
Figure 2. Top20 features for DiGeorge syndrome diagnosis prediction.

Alto: high, Bajo: low.

Figure 3. Top20 features to predict the diagnoses of SPAD (specific polysaccharide
antibody deficiency), left, and congenital agammaglobulinemia (AGAMMA), right.
Alto: high, Bajo: low.
Figure 4. Top20 features for CGD (Chronic granulomatous disease), left, and CORE
(immune deficiency, not otherwise specified), right. Notice how the features in “CORE”
suggest Job syndrome.
Alto: high, Bajo: low.

Figure 5. Top20 features for HIGM (Hyper-IgM syndrome) and SCID (Severe-combined
immune deficiency).
Alto: high, Bajo: low. Reflujo gastroesofágico: gastroesophageal reflux.
Figure 6. Top20 features for WAS (Wiskott-Aldrich syndrome), left, and NEMO (X-linked
anhidrotic ectodermal dysplasia with immunodeficiency), right.
Alto: high, Bajo: low. Leucocitos, leukocytes.

Figure 7. Top20 features for LAD (Leukocyte adhesion deficiency) with n=6, left, and
Complement deficiencies (COMPDEF) with n=12, right. Note few important features for
LAD, and a predominance of negative or absent features for complement deficiencies.
Alto: high, Bajo: low. Leucocitos, leukocytes; plaquetas, platelets.

Figure 8. Top20 of most repeated features (absence or presence) across all diagnoses. Low
IgG was included as important in 7 predictive models.
Bajo: low, Alto: high, Leucocitos: leukocytes, Reflujo gastroesofágico: gastroesophageal
reflux.

Figure 9. A flowchart of important questions to ask in the differential diagnosis of 12
classic primary immune deficiencies, derived from the SHAP values provided by our
models.
(Supplementary) list of all 238 attributes included in the model after dataset curation and
dimensionality reduction.
['Allergic rhinitis', 'Congenital heart disease', 'Cardiovascular disease', 'Autoimmune (any)',
'Heart infection', 'Constitutional symptoms', 'Bleeding', 'Palate', 'Mouth', 'Candida',
'Endocrine-Metabolic disorders', 'GI', 'IBD', 'Kidney', 'Solid tumor', 'Urinary',
'Neuropsychomotor developmental delay', 'Cancer (any)', 'Anemia', 'Arthritis', 'Joint (septic
arthritis)', 'Germs: viruses', 'CNS', 'Ear', 'Respiratory (unspecified)', 'Upper Respiratory',
'Lymph nodes', 'Flu', 'Lung', 'Angioedema', 'Eczema', 'Warts', 'Urticaria/Anaphylaxis',
'Molluscum', 'Dermatosis', 'Skin infection', 'Edema', 'Infectious cardiovascular disease',
'Serositis', 'Syndromes', 'Candida sp', 'Teeth', 'Hydroelectrolytic disorders', 'Cholecystitis',
'Jaundice', 'HepSplenomegaly', 'Food allergy', 'Diarrhea', 'Solid Tumor', 'Genital', 'Herpes',
'Proteinuria', 'Growth delay', 'Facial dysmorphism', 'Elevated blood cells',
'Hypogammaglobulinemia', 'Cytopenias', 'Endocrine Autoimmune', 'Bone', 'Bone\xa0',
'Leukemia/Lymphoma', 'Eye', 'Pneumonia', 'Congenital disorder', 'Infection (any)', 'Germs:
fungi', 'Isolates: VZV', 'Skin conditions (alopecia, vitiligo, psoriasis)', 'Nails', 'Low IgG', 'Low
IgM', 'Leukopenia', 'Cardiovascular disease, congenital', 'Bronchiectasis', 'Systemic
vasculitis', 'Cirugia_cardiovascular', 'Isolates: EBV', 'Increased inflammatory markers',
'Abscess (any)', 'Leukoplaquia', 'Sitio_boca', 'Herpes simplex', 'Sitio_oidos', 'Mumps',
'Dientes anormales', 'Sitio_senos paranasales', 'Autoinmunidad_autoanticuerpos',
'Autoantibodies', 'Absence of lymphoid tissue', 'Skin', 'Alopecia', 'Amenorrea', 'Inborn errors
of metabolism\xa0', 'Liver or viscerae', 'Autoinmunidad_hipotiroidismo', 'Sensibilidad_frio',
'Pubertad_retrasada', 'Otros_metabolicos', 'Enterovirus', 'Perinatales', 'Adenomegalia',
'Dolor_abdominal', 'Sitio_Herida_qx', 'Enzimas_hepaticas_elevadas', 'Appendicitis',
'Obesidad', 'Sitio_genitales', 'Ulceras', 'Cirugia_esplenectomia', 'Bacteria',
'Reflujo_gastroesofagico', 'Sangrado', 'Cirugia_abdominal', 'Clostrydium',

'Enfermedad_celiaca', 'Hepatitis', 'Cryptosporidia', 'Disfagia', 'Parasites', 'Poliposis',
'Colitis', 'Immune deficiency', 'Sitio_Urinaria', 'Vomito', 'Eosinofilia', 'Isolates: CMV',
'Diarrea crónica', 'Cirugia_colecistectomia', 'Giardia lamblia', 'HPylori', 'Hospitalizacion',
'Hipoalbuminemia', 'Sitio_higado', 'Malformacion_gastrointestinal', 'Hernia', 'Liver abscess',
'Falla_hepatica/cirrosis', 'Hemangioma', 'Diarrea', 'Desnutricion', 'Granulomas',
'Pseudomonas, Serratia', 'Solid tumors', 'Salmonella sp', 'Cholangitis', 'Spleen', 'Abortos',
'Malformacion_genitourinaria', 'HSV', 'Human Papillomavirus',
'Sindrome_hemolitico_uremico', 'Fistula', 'Malf_cardiovascular', 'Malf_genitourinaria',
'Autismo', 'Retraso_crecimiento', 'Alimentacion_problematica', 'Intolerancia_alimentaria',
'Perdida_peso', 'Hipoacusia_sordera', 'Blood', 'Transplant', 'Bone marrow failure', 'Elevated
antibody', 'Hemophagocyt', 'Igs_todaselevadas', 'Antibody deficiency', 'Hemofagocitosis',
'Lymph nodes\xa0', 'Monocitosis', 'Pancitopenia', 'Allergy (any)', 'Malf_gastrointestinal',
'bone', 'Dedos_anormales', 'Fracturas', 'Dolor_toracico', 'Artritis/artralgias',
'Displasia_cadera', 'SNC', 'Dolor', 'Mialgias', 'Sitio_hueso', 'Escoliosis', 'Mycobacteria',
'Alergia_respiratoria', 'Aspergillus sp', 'Sitio_articulaciones', 'Malf_esqueletica',
'Osteopenia', 'Mycobacterium sp', 'HPV', 'Linfoma', 'Coccidioides', 'Diabetes', 'Asthma',
'Fever', 'Ataxia', 'Hipertonia', 'Hipotonia', 'Oido', 'Cardiovascular desease', 'Déficit
neurológico', 'Convulsiones', 'HZV', 'Telangiectasias', 'Alimentación difícil', 'Retraso
desarrollo', 'Mycobacterium Tb', 'Acropaquia', 'Granulomas\xa0', 'Reflujo gastroesofágico',
'Environment mycobact', 'Staphylococcus', 'GramNeg', 'Pneumocystis jirovecii', 'Congenital
desease', 'Congenital cardiovascular disease', 'Hipertensión pulmonar', 'Cirugía pulmonar',
'Other infections', 'Artritis/artralgia', 'Severe viral', 'Hair', 'molluscum contagiosum',
'Systemic Lupus', 'Delayed separation of umbilical cord', 'Delayed wound healing', 'Drug
allergy', 'Necrosis', 'Skin infections', 'Eritema palmar', 'Fotosensibilidad', 'Piel redundante',
'Ulcera', 'Tejidos blandos', 'Drug Allergy', 'Piel engrosada', 'Acanthamoeba', 'Bowel
(intestines)', 'Burkholderia', 'Periodic fever', 'Catheter', 'Hpylori', 'Chronic lung disease',
'CMV/EBV', 'conjunctivitis', 'Mediastinum', 'Muscle', 'Fungemia', 'Measles', 'Streptococcus',
'Escherichia, Klebsiella', 'Other isolates', 'Mononucleosis', 'Mucosae', 'PIC line infection',
'Joints', 'Electrolyte imbalance', 'Upper respiratory', 'Urogenital/perirrectal', 'Inflammation',
'urinary', 'Gérmenes_atípicos', 'VHH8', 'Aislamiento_otros_virus', 'Gérmenes_piógenos',
'I_Bajo_EOSINOFILOS ', 'I_Bajo_IgA', 'I_Bajo_IgE', 'I_Bajo_IgG', 'I_Bajo_IgM',
'I_Bajo_LEUCOCITOS ', 'I_Bajo_LINFOCITOS', 'I_Bajo_MONOCITOS ',
'I_Bajo_NEUTROFILOS', 'I_Bajo_PLAQUETAS ', 'I_Alto_EOSINOFILOS ', 'I_Alto_IgA',
'I_Alto_IgE', 'I_Alto_IgG', 'I_Alto_IgM', 'I_Alto_LEUCOCITOS ', 'I_Alto_LINFOCITOS',
'I_Alto_MONOCITOS ', 'I_Alto_NEUTROFILOS', 'I_Alto_PLAQUETAS ']