Advanced Webinar Series on Spatiotemporal Modeling of Climate-Sensitive Diseases

Gassim · 13 January 2026 07:53

AI generated summary of the video: Introduction to spatiotemporal modeling and the webinar series

This webinar is the opening session of the Advanced Webinar Series on Spatiotemporal Modeling of Climate‑Sensitive Diseases, led by the HISP Centre (University of Oslo) together with partners like the Ministry of Health Rwanda, HISP Rwanda, CSID Network, Cardiff University’s Data Lab for Social Good, and the TRUST AI Centre. It introduces the overall process, tools, and learning path for using DHIS2‑based infrastructure and the CHAP platform for climate‑sensitive disease forecasting.

The key themes are:

Time series and spatiotemporal modeling basics:
The speaker explains time series as data observed repeatedly over time (e.g. monthly malaria cases), and “spatiotemporal” as the same type of series across many regions (districts, provinces). Models learn from historical disease data and climate covariates (e.g. rainfall, temperature) to forecast future values. Forecasts include uncertainty intervals (e.g. ranges covering 50% and 90% probability), and are evaluated by comparing past predictions against what actually happened.
Statistical underpinnings and evaluation:
The series will later cover statistical foundations such as generalized linear mixed models and Bayesian hierarchical models, along with how to choose appropriate evaluation metrics and visualizations to rigorously assess model performance.
Climate‑sensitive health impacts and early warning:
Climate change is presented as a major health threat affecting a wide range of outcomes (heat‑related illness, water‑borne disease, vector‑borne disease). Malaria and dengue are used as illustrative examples of climate‑sensitive diseases where early warning systems can support targeted interventions, stock‑preparation, and planning.
Software setup and workflows (Git/GitHub, environments, DHIS2 tools):
A central message is that a bit of extra setup effort upfront (version control with Git/GitHub, virtual environments, standardized project structure) greatly reduces the “sweating and desperation” phase of data science. The webinar outlines how climate and health data can be harmonized using the DHIS2 Climate App and Climate Tools, stored in DHIS2, and then exported as standardized CSVs for modeling.
“Hello world” modeling approach:
The speaker proposes a “hello world” style learning path: start with an intentionally simple model (e.g. a linear regression or random forest trained on a toy climate–disease dataset), but embed it in the full infrastructure – proper data files, evaluation pipeline, and a GUI front‑end – so that it already runs inside the same framework used for real cases. Once that works end‑to‑end, participants can iteratively improve the model and data.
CHAP modeling platform and DHIS2 Modeling App:
The webinar introduces CHAP (Climate Health Analytics Platform) as a reusable, open‑source platform that standardizes data input, harmonization, model training, forecasting, and rigorous evaluation (including automated generation of prediction–vs–truth plots). CHAP is tightly integrated with DHIS2 through the DHIS2 Modeling App, which lets users:
- Select models (including custom ones hosted on GitHub),
- Link them to DHIS2 data elements (e.g. dengue case counts, population, ERA5‑based climate data), and
- Run evaluations and, later, routine forecasts inside the DHIS2 ecosystem already used in 70+ countries.
Operational focus and real‑world impact:
Beyond method development, the series stresses operational use: models must match local context, be assessed rigorously on historical data, and be usable by disease programs for real decision‑making (e.g. when and where to expect outbreaks, how far ahead forecasts remain useful).
Open community and workshop in Kigali:
Participants are invited to join an open, collaborative community around CHAP and DHIS2, sharing models and evaluation approaches so countries can reuse and adapt each other’s work. The webinar also mentions a follow‑up in‑person workshop in Kigali (February 2026), with a hackathon‑style format for working on concrete use cases, for which this webinar series is a prerequisite.

The overall message is that this series will help technically prepared participants (with R/Python and some epi/stats background) learn not only how to build spatiotemporal models for climate‑sensitive diseases, but how to make them operational in DHIS2 using CHAP, with shared open‑source code, standardized workflows, and strong links to real country use cases.

AI generated summary of the video: Working with codebases

This webinar, “Working with codebases”, is the second session in the advanced DHIS2 / CHAP webinar series on spatiotemporal modeling of climate‑sensitive diseases. It focuses on setting up practical, reproducible workflows so that modeling code can be shared and reliably run by others, not just “on my machine.”

The session is structured around three pillars:

1. Terminal basics and navigation

Knut introduces the command‑line terminal as the primary interface for working with code projects in a consistent way across operating systems.

Environment setup
- macOS / Linux: Use the built‑in terminal.
- Windows: Install and use WSL (Windows Subsystem for Linux) to get a Linux terminal; this avoids later incompatibility issues.
Core navigation commands
Inside the terminal, he demonstrates how to:
- pwd – show where you are in the filesystem.
- ls – list files and folders.
- cd – change directory, including cd .. to move up.
- mkdir – create new directories for projects (e.g. a folder for this workshop).
Productivity tips
- Use Tab completion to auto‑complete long paths.
- Use the up/down arrow keys to step through command history instead of re‑typing.

The goal is to be comfortable navigating to the right project folder before running any modeling tools.

2. Version control with Git and GitHub

Next, the webinar moves to Git for version control and GitHub for sharing and collaboration.

Installing Git
- Linux: apt install git (possibly with sudo).
- macOS: brew install git (after installing Homebrew, as described in the written guide).
Cloning and forking repositories
Participants are shown how to:
- “Fork” an example repository on GitHub (Python or R version) into their own GitHub account.
- “Clone” that fork to their local machine using git clone <url>.
Basic Git workflow
Inside the cloned project, Knut demonstrates the standard cycle:
- git status – see what has changed and which files are tracked or untracked.
- git add <file> – stage new or modified files.
- git commit -m "message" – save a snapshot with a clear message (e.g. “Added new file”).
Authentication and pushing to GitHub
To avoid typing passwords repeatedly, he configures authentication with the GitHub CLI (gh), logging in via the browser once. After that, git push sends local commits to the remote repository, providing both backup and a way to share code with collaborators.

The emphasis is that Git history and good commit messages make it easy to see what changed and to roll back if a new change “breaks” the model.

3. Dependency management & virtual environments (Python and R)

The third block tackles reproducibility: ensuring others can actually run your code with the same package versions.

Python – `uv`, `pyproject.toml`, and lock files

Knut introduces uv as a modern tool to manage Python environments per project.
The project contains a pyproject.toml listing dependencies (e.g. pydantic, pandas), and uv uses this to:
- Create an isolated environment inside the project folder.
- Install exactly the needed packages and versions there.
Running code with uv run python hello.py both:
- Reads pyproject.toml / lock files to install missing libraries.
- Executes the script using that project‑specific Python interpreter.
When new packages are needed (e.g. pandas), uv add pandas updates both the environment and the lock file, so collaborators can reproduce the same setup.

R – `renv` and lock files

For R, the equivalent solution is renv:
- Install renv in the project.
- Use renv::restore() to install all required packages and versions from the project’s lock file.
This ensures that anyone checking out the project and running renv::restore() ends up with the same package versions, which is crucial when sharing models with ministries or research partners.

Both approaches directly address the common “it works on my machine” problem by making dependencies explicit and reproducible, which the presenters note is a frequent issue even for published academic modeling code.

Motivation, homework, and links to CHAP / DHIS2

Throughout, the presenters stress that these tools (terminal, Git, uv/renv) may feel like extra work at first, but they reduce manual fiddling and errors later, and are essential if CHAP models are to be reusable across countries and teams.

Participants are asked to:

Work through written guides on:
- Terminal basics (including WSL on Windows),
- Git and GitHub,
- uv for Python and renv for R.
These guides are linked from the 2026 modeling webinar series page, which is accessible via the CHAP site and the “Learning Modelling” material.

This setup is homework before the next webinar (“Introduction to the Chap modeling platform”), where participants will start running actual spatiotemporal models for climate‑sensitive diseases using these standardized workflows.

AI generated summary of the video: Introduction to the Chap modeling platform

This video is the third session in an advanced webinar series on spatiotemporal modeling of climate‑sensitive diseases, and it gives a practical introduction to the Chap (Climate Health Analytics Platform) modeling platform.
The goal is to walk technically prepared participants (with Python/R and epi/stats skills) through building and running their first forecasting model in Chap, using a complete live workflow from data preparation to evaluation.

Key steps and functionalities highlighted in the webinar include:

Getting Chap running
The session shows how Chap Core is installed and run in a local environment, so participants can execute Chap commands on their own machine or server and reproduce the examples.
Implementing a model from a minimalist example
Participants are introduced to a minimalist example model hosted on GitHub. They are shown how to copy (fork/clone) this repository so they can customize it.
The structure of this minimalist model is explained: it exposes two core functions, train and predict, which read simple CSV files, fit a model, and write out forecasts. In the basic version, a linear regression is trained to predict disease cases (e.g. dengue) from simple features such as rainfall and temperature; the presenter explicitly notes that this is intentionally oversimplified and not yet meaningful for real‑world early warning. Participants see how to run this model locally and inspect the generated predictions for a toy dataset.
The code is then briefly modified to swap the linear regression with a random forest regressor, illustrating how model developers can change algorithms while keeping the Chap interface the same.
Running models through Chap
The video explains the small configuration file (based on MLflow conventions) that tells Chap how to run the model—what command to invoke for training and for prediction, and which parameters (such as input and output file paths) are expected.
A live demonstration shows the use of the chap evaluate command, which runs the model repeatedly on a larger real dataset (e.g. public dengue data from Brazil) and generates an automatic PDF report with time‑series plots comparing predicted and observed cases.
The report clearly reveals that the minimalist model performs poorly on real data, underlining that Chap is a framework for rigorous evaluation, not a guarantee that any simple model will work.
Exploring the Chap models repository
The presenter shows that beyond the minimalist example, there is a growing repository of more advanced, open‑source models contributed by the community, including deep learning and Bayesian approaches. These are made Chap‑compatible so that they can be run and evaluated with exactly the same workflow.
Integration with the DHIS2 Modeling App
The second part of the session demonstrates how Chap models can be used through the DHIS2 Modeling App, a graphical front‑end inside DHIS2 that connects to Chap as a backend service.
From this GUI, users can:
- Select a model template exposed by Chap.
- Map model inputs (targets and covariates) to DHIS2 data elements, such as disease case counts, population, and climate variables imported via the Climate App.
- Choose time periods and organisation units for training and evaluation.
- Run evaluations without writing any code, and visualize predictions versus observed cases, including prediction intervals that show uncertainty.
Closing messages and next steps
The session closes by stressing that this “hello‑world” focused webinar is mainly about technical setup and workflow—getting participants to the point where they can run, evaluate, and iterate on models in Chap and DHIS2. Once this pipeline is in place, modelers can collaborate with public health experts to develop highly tailored forecasting models and share them as open resources for other countries
Subsequent webinars in the series are announced as covering more conceptual topics: climate factors and health impacts, statistical modeling foundations, and advanced methods for spatiotemporal disease forecasting.

AI generated summary of the video: Webinar: Climate factors, climate change, and their health impacts

This webinar in the Spatiotemporal Modeling of Climate‑Sensitive Diseases series introduces how climate factors, climate change, and non‑climatic drivers interact to affect health, with a focus on climate‑sensitive infectious diseases and their use in early warning systems.
Two researchers from the Barcelona Supercomputing Center use dengue as a recurring case study to illustrate concepts.

Climate‑sensitive diseases and hazards
The presenters first explain how climate variables such as temperature, precipitation, and humidity give rise to hazards (heatwaves, floods, droughts), which can impact health directly (e.g. heat stress, injuries) or indirectly through malnutrition and diarrhoeal diseases.
They then focus on climate‑sensitive infectious diseases:

Vector‑borne diseases like malaria, dengue, chikungunya, where vectors are highly sensitive to temperature, moisture, and water availability.
Water‑borne diseases like cholera and leptospirosis, where flooding, sea‑level rise, and drought affect water and sanitation systems.
These diseases typically show strong seasonality, distinct geographic ranges, and multi‑year variability.

Climate drivers and disease patterns
The webinar explores how specific climate conditions shape risk. For example, drought can increase dengue risk when people store water in containers that favour Aedes aegypti breeding, whereas very intense rainfall can flush out larvae and reduce risk.
Large‑scale ocean phenomena such as El Niño and La Niña are discussed as drivers of temperature and rainfall anomalies that propagate into disease patterns.

Key properties of climate–disease relationships
Three recurring properties are highlighted:

Non‑linearity: relationships are often hump‑shaped—both “too little” and “too much” rain or heat can reduce transmission by disrupting vectors or pathogens.
Lags: there is usually a delay between climate anomalies and observed cases, driven by vector life cycles, pathogen incubation, and human behaviour.
Modulation by non‑climatic factors: socioeconomic conditions, demographics, land use, water/sanitation infrastructure, and behaviour can strengthen or weaken climate effects, and interact with each other and with climate in complex ways.

Climate‑informed early warning systems (EWS)
The second half of the webinar shows how these insights are translated into climate‑informed EWS that combine climate and epidemiological data to forecast outbreaks and support more timely, targeted responses. Typical lead times for vector‑borne disease EWS are on the order of 1–6 months, to allow time for interventions. Building such systems involves:

Stakeholder engagement and problem definition (disease, lead time, spatial scale).
Assessing evidence on relevant climate drivers.
Assembling and harmonizing data (climate, environmental, demographic, surveillance).
Model development and validation against historical data, explicitly checking how well predictions would have worked in the past.
Risk communication and co‑definition of actions linked to different alert levels (e.g. enhanced surveillance, vector control, community communication, facility preparedness).
Continuous evaluation and iteration in real use.

Modeling approaches for EWS
Three families of models are compared:

Statistical models (e.g. time‑series regression) that relate cases to predictors under probabilistic assumptions, work well with long historical series, and give principled uncertainty estimates.
Machine‑learning models (e.g. random forests, neural networks), which are flexible and potentially more accurate but often require more data and are harder to interpret and to use for uncertainty quantification.
Mechanistic models (e.g. compartmental or agent‑based models), which encode biological and behavioural processes such as immunity, contact patterns, and mobility; these can model scenarios but depend on strong assumptions and detailed system knowledge.

Data needs and preprocessing
The presenters outline four main data categories needed for climate‑informed EWS:

Weather and climate data from national meteorological stations and/or global reanalysis products.
Environmental data, including land use/cover, vegetation indices (e.g. NDVI), and water indices derived from remote sensing.
Demographic and socioeconomic data to represent population at risk and vulnerability.
Disease surveillance data, usually case counts reported at weekly or monthly resolution.

They then describe three core preprocessing strategies:

Rolling statistics: computing moving averages (e.g. 3–12‑month temperature means) or moving sums of precipitation to capture persistent warm/cold periods, droughts, or extended wet spells.
Anomalies: expressing current climate values relative to a long‑term baseline (e.g. 30‑year mean and variance) to highlight departures from “normal” conditions and identify extremes.
Lags: explicitly including delayed versions of climate variables (e.g. rainfall 1–6 months ago) so models can predict future risk using observed past climate, reducing dependence on uncertain forecasts.

The webinar closes by linking these methodological pieces back to the broader series and to tools like Chap, which aim to make rigorous climate‑informed modeling and evaluation more accessible to public‑health practitioners and modelers working with DHIS2 data.

AI generated summary of the video: Webinar: Fundamentals of statistical modeling

This webinar on Fundamentals of Statistical Modeling is part of an advanced series on spatiotemporal modeling of climate‑sensitive diseases. It aims to equip technically prepared participants with core statistical concepts needed to build and evaluate forecasting models for diseases such as dengue or malaria, especially in the context of early warning systems.
The session focuses on how to move from simple linear models to more appropriate spatiotemporal models for disease case data, and on understanding and communicating uncertainty in predictions.

The basic steps of modeling and forecasting with time‑series data: preparing data, identifying patterns such as trends and seasonality, fitting models, and generating predictions for climate‑sensitive diseases.
Core ideas from Bayesian statistics, emphasizing how Bayesian methods provide full probability distributions for parameters and predictions, and why this matters for operational early warning systems where uncertainty must be quantified and communicated.
Different sources of uncertainty in predictive modeling: variability in the data, limitations of model structure and assumptions, and external climate or environmental drivers that affect forecast accuracy.

AI Summary based on the detailed transcript‑derived notes, the webinar walks through the following technical elements:

Linear model review: Starting from a simple linear regression where disease case counts are modeled as a function of rainfall and temperature, the presenters discuss assumptions such as linearity, conditional independence, and normally distributed errors, and why these are problematic for count data.
Generalized Linear Models (GLMs): GLMs are introduced as a way to handle non‑normal outcomes and non‑linear relationships, using appropriate distributions for counts (e.g. Poisson or negative binomial) and link functions (such as the log link) to connect the mean response to a linear predictor and to interpret coefficients on the link scale.
Population offset: To move from crude counts to rates, the model includes a population offset, allowing comparisons across spatial units and times with different population sizes by effectively modeling incidence rates.
Bayesian inference in practice: Beyond the conceptual introduction, the session shows how Bayesian inference is applied to these GLMs to obtain posterior distributions, credible intervals, and probabilities that predicted values exceed defined thresholds—key quantities for triggering early warning alerts.
Hierarchical (multilevel) models: Hierarchical structures are used to incorporate group‑level information (e.g. regions or districts) while retaining unit‑level detail (e.g. municipalities). Partial pooling via varying intercepts and slopes shares information across groups and stabilizes estimates where data are sparse.
Model terms for spatiotemporal structure: The model is extended with terms to capture interannual variation and seasonality (e.g. annual random effects, monthly seasonality components with suitable priors) and to model spatial dependence (e.g. intrinsic CAR/BYM2‑type structures) so that neighbouring areas can borrow strength from each other.
Non‑linear effects: To represent non‑linear climate–disease relationships, the webinar illustrates the use of polynomial terms and/or spline functions on climatic covariates, increasing model flexibility while retaining interpretability.
Forecasting for early warning systems: The full predictive distribution from the Bayesian model is then used to frame early warning in probabilistic terms—for example, computing the probability that incidence will exceed an outbreak threshold, and using this for risk communication and decision rules.
Model evaluation: Finally, the webinar emphasizes the role of cross‑validation and retrospective evaluation—comparing out‑of‑sample forecasts with observed outcomes over historical periods—to assess and compare model performance before operational use.

AI generated summary of the video: Webinar: Using DHIS2 Climate Tools - A flexible toolkit for climate data integration

This webinar presents DHIS2 Climate Tools, an open‑source Python toolkit for accessing, processing, harmonizing, and loading climate, weather, and environmental data into DHIS2 and the Chap modeling platform. The tools are aimed at data engineers and data scientists who need reproducible workflows for exploring, analyzing, visualizing, and importing both global and local data sources.

Introduction to DHIS2 Climate Tools
The webinar starts with an overview of the toolkit and its website entry point, which documents open‑source libraries and end‑to‑end workflows for integrating climate, weather, and environmental data with DHIS2 and Chap.
Climate Tools is built on Python and leverages the geospatial ecosystem (Xarray, GeoPandas, EarthKit) to make complex spatial operations manageable.

Challenges and solutions in climate data integration
Common challenges described include accessing climate data from different providers and aligning it with health data in space and time. To address this, the toolkit provides dhis2eo, a Python library that can download climate and environmental datasets (e.g. from the Copernicus Climate Data Store or the Climate Hazards Center) and convert them towards DHIS2‑ready formats.

Data harmonization (temporal and spatial)
A central part of the webinar is GIS‑based harmonization:

Temporal harmonization: many global datasets (e.g. ERA5) are available hourly. Climate Tools shows how to aggregate these to daily or other DHIS2‑relevant periods, handling time aggregation as part of the workflow.
Spatial harmonization: global products are typically gridded (e.g. ~9 km cells). Using Xarray, GeoPandas, and EarthKit, the toolkit aggregates grid cells to DHIS2 organisation units by overlaying the grid with admin polygons and computing statistics (e.g. mean temperature or rainfall per district).

These steps produce datasets directly aligned with the health regions where DHIS2 data are collected.

Automating data flows and integration options
The webinar emphasizes that Climate Tools is meant for automated, repeatable pipelines: once a workflow is defined, it can be run regularly (e.g. as a scheduled job) so DHIS2 always has up‑to‑date climate and environmental data, which is particularly important for early warning and forecasting use cases.
The outputs can be:

Imported into DHIS2 (as aggregate data values), or
Sent directly to the Chap modeling platform when working primarily on modeling workflows rather than DHIS2 storage.

Python client for DHIS2 interaction
A key component is the DHIS2 Python Client, bundled in the Climate Tools ecosystem. It provides a lightweight interface to the DHIS2 Web API from Python, making it easier to pull and push data, work with metadata, and integrate climate workflows into broader data engineering tasks.

Contributions and community workflows
Finally, the webinar invites contributions from the community. The project maintains a growing set of runnable Jupyter notebooks that demonstrate end‑to‑end workflows (downloading, harmonizing, and uploading data), and users are encouraged to contribute their own notebooks and use cases to the GitHub repositories, making Climate Tools a collaborative, evolving toolkit.