Portfolio

Bayesian Causal Mediation between Ancestry and Lung Function (2025)

I sought to estimate causal link between race-neutral lung function and global genetic ancestry (GGA). RWD (ED visits) were extracted using ICD9/10 billing codes from Cincinnati Children's EHR system. Only patients with GGA and lung function information, available during the ED encounter were included.
Highlights:

Raw FEV1 values were adjusted via race-neutral reference equation.
Global genetic ancestry per patient was obtained using RFMix.
Bayesian structural equation model consisted of five explicit sub-models with a random effect using brms.
Causal ADEs, ACMEs, and % of contribution (mediation) of GGA variation on FEV1 were estimated.

Code (GitHub) -> URLPaper -> Tbd
#epidemiology #precisionmedicine #EHR #mediation #bayesian #sem

Fairness in ML: Detecting Algorithm Bias in Asthma Models

(2025)

This project aimed to detect and mitigate algorithmic bias between race groups in our previously developed models for asthma hospitalization in US adults. The dataset included asthma visits identified using ICD9/10 billing codes, 22 risk factors and two health outcomes.
Highlights:

Diagnostic and fairness metrics were calculated across all race/ethnicity groups (AUC, sensitivity, specificity, equal opportunity rate, disparate impact).
T-tests on boostrap simulations were used to assess differences between White and non-White predictions.
Mitigation strategies included several data (oversampling) and modeling strategies (tree-based approaches, predictor screening) to reduce unfair predictions.

Code (GitHub) -> URLPaper -> Tbd
#epidemiology #machinelearning #precisionmedicine #EHR #fairness

Bayesian Event-free Time and Time to Recurrence of Asthma Patients in ED

(2024)

This project aimed to predict the probability of ED encounter-free survival time and time to recurrence (TTR) of ED encounter among asthmatic children with allergic comorbidities. ED visits were identified using ICD9/10 billing codes from Cincinnati Children's Hospital. Causal covariate selection and Bayesian survival analyses were used to assess relationships between mediators, and number of ED visits.
Highlights:

Causal backdoor criterion was used to remove open paths of longitudinal measurements of ED encounter characteristics (RWD EHR data).
Jaccard index was estimated to measure similarity between longitudinal comorbid diagnoses at ED encounters.
Bayesian mixed-effects survival analysis with 10 degrees M-spline survival function was implemented for modeling hazard ratios (HRs).
Time to recurrence (TTR) was estimated using the Restricted Mean Survival Time (RMST).
Validation used the expected log-predictive density (ELPD) with leave-one-group-out (LOGO) scheme to compare M-spline against exponential models. Calibration and Pareto Smoothed Importance Sampling (PSIS) plots were provided.

Code (GitHub) -> URLNotebook tutorial -> URLPaper -> URL
#epidemiology #precisionmedicine #EHR #survivalanalysis #bayesian

Assessing Length of Stay and Readmission for Asthma in Adults using Bayesian LMMs (2024)

This project aimed to predict asthma length of stay and readmission among adults in US. In-patient visits were identified using ICD9/10 codes from All of US cohort (EHR and survey data). Causal covariate selection and Bayesian mixed effects models were used to predict outcomes.
Highlights:

Performed data extraction using python and BigQuery in the AoU's google cloud environment.
Use of a directed acyclic graph (DAG) to aid causal covariate adjustment.
Employed Bayesian mixed-effects negative binomial and Logistic regression models to estimate the LoS in days and readmissions within a 30-day period.
Accounted for subject-specific and multi-site variance including random intercepts for patients with recurrent asthma-related visits, and for EHR sites.
Cross-validated using Bayesian ELPD and leave-one-group- out (LOGO) scheme.

Code (GitHub) -> URLPaper -> URL
#epidemiology #machinelearning #precisionmedicine #EHR #fairness

Predicting COVID19 Hospitalization of Children with Allergies (2023)

This project aimed to predict hospital utilization of children with allergies (no hospital care, short stay or prolonged stay). In-patient visits were identified using ICD9/10 billing codes from Cincinnati Children's Hospital during COVID19. Causal mediation and mixed-effects analyses were used to assess relationships between mediators, and number of ED visits.
Highlights:

Causal framework was used for proper covariate adjustment.
Causal mediation analysis for direct and indirect effects quantification on COVID19 hospitalization.
Mixed-effects LogReg model was implemented for modeling RWD (EHR admission discharge visits).
Cubic splines smoothing were required to model non-linear and stationary periodic effects when required.
Validation was performed in both training and testing data using AUROC and AUPRC.
Disease mapping in R was performed for geospatial interpretability.

Code (GitHub) -> URLPaper -> URL
#epidemiology #geospatial #precisionmedicine #EHR #splines #causalinference

Time-Series Forecasting of ED Utilization During COVID19 (2023)

This project aimed to compare average weekly ED encounters during the pre- and intra-COVID-19 periods. ED visits were identified using ICD9/10 billing codes from Cincinnati Children's Hospital during COVID19. I implemented an interrupted time-series models to evaluate weekly ED rates (per 10,000 population) by race and diagnosis.
These models allow the estimation of changes in rates over time (trends) and seasonality before and during an event is introduced (aka counterfactual; eg, the national emergency declaration of COVID-19)
Highlights:

Use of interrupted time-series (GAMs) for anomaly detection.
Cubic splines restricted to 53 knots (number of weeks) to model seasonality in ED visits.

Code (GitHub) -> URLPaper -> URL
#epidemiology #timeseries #precisionmedicine #EHR #splines

Predicting Asthma ED Visits in Cincinnati, OH (2022)

This project aimed to predict number of asthma-related ED visits per year among children. Ten year of ED visits were identified using ICD9/10 billing codes from Cincinnati Children's Hospital. Causal mediation and mixed-effects analyses were used to assess relationships between mediators, and number of ED visits.
Highlights:

Causal framework was used for proper covariate adjustment.
Causal mediation analysis for direct and indirect effects quantification of risk factors.
Mixed-effects negative binomial model was implemented for modeling RWD (EHR recurrent ED visits).
Cubic splines smoothing were required to model non-linear and stationary periodic effects when required.
Validation was performed in both training and testing data through mean absolute error (MAE).
Disease mapping in R was performed for geospatial interpretability.

Code (GitHub) -> URLPaper -> URL
#epidemiology #geospatial #precisionmedicine #EHR #splines #causalinference

Small Area Estimation (SAE) & Spatial Autocorrelation using R-INLA (2020)

As part of the COVID-19 Task Force in Ohio, I implemented a Bayesian hierarchical spatial model with an Intrinsic Conditional Auto-Regressive (ICAR) to estimate the risk of COVID-19 related death per county. Counts of COVID-19-related deaths in the U.S. were obtained from the famous Johns Hopkins University database for 3142 counties from all 50 states.
Highlights:

ICAR model with adjacency matrix was used to estimate COVID19 mortality risk and spatial effect (Φ) for connected and disconnected territories.
Nested county-state structure was induced for estimation of random effects.
Spatial clustering was evaluated using Moran's I statistic test on residuals.
Bivariate disease mapping was performed for geospatial interpretability.

Paper -> URL
#epidemiology #geospatial #bayesian #covid19

Web-based EEG Viewer and Tooler for Alzheimer’s Records (2018)

UBT-Alzheimer is a neuroimaging web-tool developed for the National Department of Science - Colciencias through the grant #744 "CONVOCATORIA PARA PROYECTOS DE CIENCIA, TECNOLOGÍA E INNOVACIÓN EN SALUD 2016". It allows visualization of patient's EEG recordings, modular signal processing (notch and ocular filters) and neural activation via 3D mesh (Loreta). Experts can load EDF files and navigate through their metadata info with intuitive UI paradigms. Three.js, d3js were used for the 2D/3D Viz, while front and backend were developed in Angular and NodeJS.Code (GitHub) -> URL
#medicalimaging #dataviz

Semi-automatic Coronary Centerline Extraction Framework (2013)

As part of my master studies, I aimed to implement a semi-automatic method for coronary artery centerline extraction from computed tomography angiography data. Using two seed points (start and end) provided by a radiologist, the centerline was extracted interactively via a medialness image [1][2] and a minimum cost path algorithm [3]. Developed entirely in C++ with ITK, VTK, and the CreaTools framework [4], current work focuses on comparing additional medialness techniques.References:
[1] E. Correa-Agudelo et al., A Modular Workflow Architecture for Coronary Centerline Extraction, ICCVG 2014. (PDF)
[2] CreaTools: http://www.creatis.insa-lyon.fr/site/en/CreaTools_home
[3] Master Thesis: Chapter 3#computervision #dataviz

cvTraffic: A Computer Vision Based Tool for Traffic Analysis (2013)

This project aimed to develop a computer vision framework for traffic analysis. Vehicle counting and tracking utilized MoG segmentation [1] and bounding box overlap heuristics. Classification employed BoW+SVM [2] to distinguish vehicle classes. Statistics were accessible via a publisher-subscriber API (zeroMQ) for offline/online use. Current execution time is 33–34fps on fanless microITX.References:
[1] Z. Zivkovic & F. van der Heijden, Efficient Adaptive Density Estimation for Background Subtraction, Pattern Recognition Letters, 2006.
[2] G. Csurka et al., Visual Categorization with Bags of Keypoints, 2004.[3] http://sirius.utp.edu.co
#computervision

Basics of tracking: Pedestrian Tracking using camshift in OpenCV (2012)

Pedestrian tracking is vital for video surveillance. I combined Comaniciu's mean shift algorithm with the Kalman filter [2]. Frames were converted to the CIE1976 (Lab*) color space to distinguish targets. The adaptive mean shift algorithm [3] calculated window size and position, while the Kalman filter handled occlusions. Experiments showed this approach outperformed using Camshift alone.
#computervision

HCI - Two-Handed Interaction in Virtual Environments (2012)

Our interest in this project is to show the development of Mixed Reality experiences, in which interaction with novel input and output devices should be taken into account. However, in order to integrate hardware solutions, we require software technologies that hide part of their complexity and facilitate development. We implement a common two-handed 3D interaction to show some results from our preliminary user studies and lessons learned. All the code is developed in java and the connection to the devices is handled by a VRPN server.Paper -> RG
#hci

Stereo imaging using XNA + Kinect (2011)

At the end of 2011, I teamed-up with researchers from Uniandes visual computing group Imagine (Colombia) and local mobile network operator (Tigo) to develop an interactive kiosk for advertising purposes. In the course of this project, I programmed stereoscopic visualization in the 3D engine (XNA). Also, I was involved in the semiotics of the user interface and gesture interaction with novel devices (Microsoft Kinect sensor and Phasespace Motion Capture).
#hci #virtualreality

iPuppet: Interactive Puppet Show (ACE-2011)

We proposed an interactive puppet show via tele-Operated system where kids can be part of the story. This solution enables a new interaction component in the show using a Microsoft Kinect camera, and a puppet theatre with a rear projection screen. We followed a User-Centered Design process, with about 20 kids that identified the requirements and possibilities of this new interface.Paper -> RG
#hci

Color-based face detector (2009)

Oldy but goldy RGB-based face detection in OpenCV (C++).
#computervision

Multitouch screen using OpenCV (2009)

What a time to learn OpenCV 1.1 via this cool homemade project mimicking a multitouch screen.
#computervision #hci

Google Sites

Report abuse