Ghana Tuberculosis Forecasting — Christopher Odoom

01 / The question

How might Ghana’s TB burden change by 2030?

The historical series is short—34 annual observations—so a more complex model is not automatically a better model.

The analysis uses IHME Global Burden of Disease 2023 estimates to describe long-run changes and test whether neural or hybrid models improve six-year holdout prediction over transparent time-series baselines.

Study design

Train on 1990–2017, evaluate on 2018–2023, select a model separately for each endpoint, refit on the complete series, and forecast 2024–2030.

02 / Methodology

Eight models, one honest holdout period.

01

Prepare endpoints

Build annual age-standardized incidence, mortality, and mortality-to-incidence ratio series.

02

Diagnose trends

Use ADF/KPSS tests, ACF/PACF plots, rolling statistics, and endpoint-specific differencing.

03

Fit candidates

Compare Naive, Drift, ARIMA, ETS, LSTM, Causal TCN, ARIMA–LSTM, and multistep LSTM.

04

Evaluate

Rank 2018–2023 predictions using MAE, RMSE, and MAPE, without training on the test period.

05

Propagate uncertainty

Simulate historical GBD trajectories and combine input uncertainty with empirical holdout residuals.

Historical tuberculosis incidence, mortality, and mortality-to-incidence ratio in Ghana — Primary age-standardized TB series used for modeling, 1990–2023.

03 / Findings

Incidence and mortality fell sharply—but different models won.

Incidence, 1990–2023−59.9%

485.2 to 194.7 per 100,000

Mortality, 1990–2023−67.7%

125.5 to 40.6 per 100,000

2030 incidence forecast146.5

LSTM point forecast per 100,000

The best holdout model depended on the endpoint. An LSTM gave the lowest error for incidence, while a simple linear drift model performed best for mortality and ETS performed best for the mortality-to-incidence ratio. That result is a useful reminder that small annual datasets often favor simpler models.

Selected models on the 2018–2023 holdout
Endpoint	Selected model	RMSE	MAPE	2030 point forecast
Incidence rate	LSTM	9.544	3.67%	146.51
Mortality rate	Drift	0.644	1.14%	22.59
Mortality-to-incidence ratio	ETS	0.0095	4.14%	0.216

04 / Model comparison

Evaluation before extrapolation.

Error metrics for eight forecasting models across three TB endpoints — Holdout error comparison across all classical, neural, and hybrid candidates.

Best-model forecasts for Ghana TB endpoints through 2030 with uncertainty bands — Selected-model forecasts through 2030. Bands distinguish GBD input uncertainty and combined input-plus-model uncertainty.

05 / Interpretation

Forecasts are scenarios, not guarantees.

Thirty-four observations constrain every model.

GBD values are modeled estimates rather than direct surveillance counts, and structural changes in diagnosis, reporting, treatment, or policy could break historical patterns. Neural results are also sensitive to short-series training choices. The repository reports uncertainty and retains simple baselines so the additional complexity can be judged rather than assumed.