SENTINARA

Glossary

The words behind the charts: what each measure means, how each signal is computed and why it matters for early warning.

Key facts

  • Dengue: A viral infection caused by four related dengue viruses (DENV-1 to DENV-4) and spread by Aedes mosquitoes. Most infections are mild, but a small share progress to severe dengue with plasma leakage, bleeding or organ impairment.
  • Dengue haemorrhagic fever (DHF): A severe form of dengue marked by plasma leakage and a low platelet count. Indonesia reports DBD separately from milder dengue fever, which is why Semarang data has two series.
  • Dengue fever (DF): Dengue without plasma leakage: high fever, headache, pain behind the eyes, muscle and joint pain, sometimes a rash.
  • Severe dengue: The WHO 2009 category for dengue with severe plasma leakage (shock or fluid accumulation with breathing difficulty), severe bleeding or severe organ involvement. It needs hospital care.
  • Serotype: One of the four dengue virus types. Infection gives lasting immunity to that type only; a later infection with a different serotype carries a higher risk of severe disease. A change in the dominant serotype often precedes large outbreaks.
  • Aedes aegypti and Aedes albopictus: The mosquitoes that transmit dengue. They breed in clean standing water around homes and bite mainly during the day, especially in the morning and late afternoon.
  • Critical phase: Usually days 3 to 7 of illness, often when the fever falls. Plasma leakage can start, so warning signs must be watched even if the patient seems to improve.
  • PSN 3M Plus: Indonesia's mosquito-nest eradication programme: drain and scrub water containers (Menguras), cover them (Menutup) and recycle items that collect water (Mendaur ulang), plus extra measures such as repellent, bed nets and larva-eating fish.
  • Larva-free index (ABJ): The percentage of inspected houses where no mosquito larvae are found. Indonesia's programme target is at least 95%.
  • Wolbachia: A natural bacterium that, when introduced into Aedes aegypti, reduces the mosquito's ability to transmit dengue. Mosquitoes carrying Wolbachia are released to spread it through the local population.
  • Incidence rate (IR): New cases per 100,000 people in a year. It makes places with different population sizes comparable.
  • Case fatality rate (CFR): Deaths divided by reported cases, as a percentage. It reflects both disease severity and the quality of clinical care and reporting.
  • Seasonality: The regular rise and fall of cases through the year. In Indonesia most cases occur from January to April, during and after the rainy season.
  • Year to date (YTD): Cases from January to the latest complete month. Comparisons always use the same months of earlier years.
  • Latest complete month: The most recent month considered fully reported. The current month and trailing months far below the recent level (late reports) are excluded.
  • Endemic channel (Bortman): For each calendar month, the median, 3rd quartile and epidemic threshold of the five previous complete years. The current month is placed in one of four zones. Widely used by PAHO and WHO.
  • Epidemic threshold: The upper line of the endemic channel: the geometric mean plus 1.96 standard deviations of log(cases + 1) for that month in the previous five years.
  • Success, Safety, Alert and Epidemic zones: The four bands of the endemic channel: below the median (Success), median to 3rd quartile (Safety), 3rd quartile to the epidemic threshold (Alert) and above the threshold (Epidemic).
  • Warning level: SENTINARA's overall assessment for a place: Normal, Watch, Alert or Outbreak signal. It combines the endemic-channel zone, the outbreak score, CUSUM, the forecast threshold risk and momentum, and lists the reasons behind it.
  • Reporting reliability: How far a reporting stream can be trusted for alerts. It is low when reports arrive in batches or with long runs of zeros; such places can be raised to Watch at most.
  • Momentum: The change between the average of the last three months and the three months before. Rapid momentum above the median is an early sign of a season starting.
  • Outbreak score (Farrington-style): Compares the current month with the same month and its neighbours in the previous five years, after a 2/3-power transformation. A score above 1 means cases exceed the 99.5% upper bound expected for that time of year.
  • Robust z-score: How unusual a value is after removing trend and seasonality, measured with the median and the median absolute deviation (MAD) so that past outbreaks do not hide new ones.
  • CUSUM: Cumulative sum control chart. It adds up how far each month is above normal and raises an alarm (at 4 or more) when small excesses persist, catching slow, sustained rises that single-month tests miss.
  • Forecast ensemble: A weighted average of three models: seasonal naive, damped Holt-Winters and harmonic regression with autoregressive terms. Weights come from each model's backtest error, and much weaker models get zero weight.
  • Holt-Winters (damped): Exponential smoothing that tracks a level, a trend and a 12-month seasonal pattern. The trend is damped so long-range forecasts do not run away.
  • Harmonic regression with AR terms: A ridge regression on log(cases + 1) using sine and cosine terms for the yearly cycle plus the values of 1, 2 and 12 months earlier.
  • Rolling-origin backtest: Repeatedly pretending it is an earlier month, forecasting forward with only the data available then, and comparing with what actually happened. It measures real forecasting skill rather than fit.
  • MAPE: Mean absolute percentage error of one-month-ahead forecasts in the backtest. It is easy to read but inflated for places with very small counts.
  • MASE: Mean absolute scaled error: forecast error divided by the error of simply repeating last year's value. Below 1 means the model beats that naive benchmark.
  • Prediction interval and coverage: The range the actual value is expected to fall in (80% or 95%). SENTINARA builds it from the ensemble's own backtest errors and reports coverage, the share of past months that really fell inside.
  • Threshold risk: The probability, from the forecast distribution, that at least one of the next three months exceeds its epidemic threshold. Used only when the forecast has shown skill (MASE ≤ 1).
  • Seasonal similarity: Cosine similarity between two places' monthly share of yearly cases. Places with similar values peak in the same months.
  • WHO Global Dengue Surveillance: The World Health Organization's public database of dengue cases and deaths reported by member states, available through an open API. SENTINARA's main source for current data.
  • OpenDengue: A peer-reviewed open database (Clarke et al., 2025) that harmonises national and sub-national dengue reports since the 1920s. Used for long-run history and Indonesian provinces.
  • Series splicing: Extending a WHO series backwards with OpenDengue months. It is done only when both sources agree on the months they share (median ratio between 0.67 and 1.5).
  • Negative correction: Some countries publish negative weekly counts to correct earlier cumulative totals. SENTINARA absorbs them into the preceding months of the same year, so no month is negative and the annual total is unchanged.
  • Data validation: Automatic checks run on every refresh: shapes and counts, no future months, annual totals equal monthly sums, sources agree, forecasts are valid. A failed check stops the update before it is published.
  • SHA-256 checksum: A fingerprint of each data file listed in the manifest. Anyone can recompute it to verify that a downloaded file is exactly what SENTINARA published.

Data updated 2026-10-31. Sources: WHO Global Dengue Surveillance; OpenDengue V1.3 (Clarke et al., 2025), doi:10.6084/m9.figshare.24259573; Dinas Kesehatan Kota Semarang; World Bank Open Data; NASA POWER.