Table 3. Main findings, conclusions, and limitations of the studies.
| Article ID | Main study findings, including prediction quality (sensitivity + PPV) | Temporal and spatial risk prediction | Prediction time lag | Study or model limitations | Conclusions |
|---|---|---|---|---|---|
| D1 [21] | Improved prediction, user-friendly, implementable tool. Sensitivity: 50%–100% and PPV: 40%–88% | High temporal prediction, low spatial prediction (district level) | A range of 1–12 weeks | Lacks predictions at small spatial unit and requires weekly to enable operational forecasting | The tool is pragmatic and useful for detecting imminent outbreaks |
| D2 [26] | Operationally useful, LASSO was superior to methods (SARIMA model) except the first 2-week window | High temporal precision, low spatial prediction (district level) | 12 weeks | LASSO methods are not amenable to interpretation (mainly at longer forecasting window), hindered by the numerous covariates acting at different lags | Automated machine learning methods such as LASSO can markedly improve forecasting techniques |
| D3 [12] | Sensitivities: 72%–97% and PPV: 45%–86% at a lag of 1–12 weeks | High temporal precision, low spatial prediction (district level) | 1–12 weeks | Can be disturbed by inconsistent and missing data especially with regard to entomological indices | Probable cases and meteorological variables indicate for increased risk of transmission |
| D4 [31] | Sensitivity: 50%–100% and specificity: 75%–100% | High temporal precision, low spatial prediction (district level) | 5 weeks | Algorithm used needs to be trained, which may cause a loss of robustness if the outbreak pattern changes or differs significantly from previous years | Surveillance R-package algorithms are free and implementable. Time-space trends monitoring can also be useful |
| D5 [28] | Sensitivity: 100% and specificity: 99.8% and a median time to detection of 3 days | Low temporal prediction but high spatial prediction | 3 days | N/A | CIDARS had good sensitivity, specificity and timeliness of outbreak detection |
| D6 [27] | AUCs are 75% for forecasting 12 weeks and 80% for 5 weeks in advance | High temporal and high spatial predictions | 1–12 weeks | The model is highly reliant on a rich dataset of georeferenced case identifications and demand regular update and the adaptation require pre-adjustments to the grid used in different geo-areas. | Spatially resolved forecasts of geographically structured diseases can be obtained at a neighborhood level in urban and rural environments for guiding control efforts |
| D7 [57] | A combination of surveillance and meteorological was optimal; temperature at lag 3 weeks, rainfall at lag 2 weeks, and rainfall at lag 3 weeks. Sensitivity: 88.9%, specificity: 81.0%, PPV: 74.4%, and NPV: 92.2% | High temporal level but low spatial prediction | 12 weeks | Predictive model could explain only 64% of the variation in the occurrence of cases and is biased by underreporting of cases | Past disease incidence data, up to years, are crucial predictive possibly indicating cross-immunity status of the population |
| D8 [30] | Models for describing, simulating, and predicting spatial patterns of Aedes aegypti populations associated with climate variability patterns | Unknown temporal but low spatial prediction | N/A | N/A | Using indices of climate variability can construct spatial models providing warning of potential changes in vector populations in rainy and dry seasons and by months |
| D9 [38] | Sensitivity: 75% (lag, 1–5 months) and PPV: 12.5%. Climate predictors were good classifiers of risk areas based on the different climate in different regions | High temporal and high spatial predictions | 4–52 weeks | The model is limited to issuing alerts with short-time intervals (1–5 months ahead), which may not be practical in operational modes | It is possible to detect dengue outbreaks ahead of time and identify populations at high risk |
| D10 [23] | A 9% increase in the incidence of imported cases for every additional 10,000 travelers arriving from affected areas | No temporal prediction but low spatial prediction | N/A | N/A | The risk of disease importation was computed with the volume of international traveler from disease-affected areas worldwide |
| D11 [43] | Time series Poisson model using climate data well predicted at time lag of 3 months after controlling the autocorrelation, seasonality, and long-term trend | High temporal but low spatial prediction | 48–96 weeks | N/A | Transmission vector Aedes albopictus, imported cases, monthly climatic information are useful for cheap and effective EWS |
| D12 [41] | Sensitivity/specificity: 78%/92% for a threshold of 3 cases per week, sensitivity/specificity: 91%/91% for a threshold of 2 cases per week, and sensitivity/specificity: 85%/87% for a threshold of 1 case per week | Low temporal prediction but no spatial prediction | 1 week | Limited to climatic factors and can be biased by underreporting of cases | Occurrence of outbreaks in the study city could impact disease outbreaks in neighboring city under suitable weather conditions |
| D13 [42] | Models with DBSI (ICC: 0.94 and RMSE: 59.86) is better than the model without (ICC: 0.72 and RMSE: 203.29) | Low temporal but poor or no spatial prediction | 1 week | Uses short-term time series data and prone to confounding effect | DBSI combined with traditional disease surveillance and meteorological data can improve the dengue EWS |
| D14 [44] | Summer Equatorial Pacific Ocean sea surface temperatures and Azores high sea-level pressure model correctly predicted 80% and missed 15% of the nonepidemic years | Low temporal and Low spatial prediction | Annual (year-to-year variability) | N/A | Outbreak resurgence can be modeled using a simple combination of climate indicators |
| D15 [40] | Environment-based, multivariate, autoregressive models predicted 2–26 weeks ahead | High temporal and Low or no spatial prediction | 8–26 weeks | N/A | Outbreaks often occurred when extreme daily temperatures are confined within the 18–32°C range, Patterns of spatial variability across endemic regions may be related to variations in the built environment, ecology, local weather, population density, mitigation efforts, and host mobility |
| D16 [45] | SVR model selected by a cross-validation technique accurately forecasted at 12 weeks with smallest prediction error | High temporal but low or no spatial prediction | 12 weeks | Internet searching behavior is susceptible to the impact of media reports, which may affect the performance of the model | SVR model achieved a superior performance in comparison with other forecasting techniques |
| D17 [13] | The model predicted accurately with <3% false alarm | Good temporal but poor or no spatial prediction | 16 weeks | N/A | Models using temperature and rainfall could be simple, precise, and low-cost tools for disease forecasting |
| D18 [35] | SARIMA model is robust and autoregression, moving average and seasonal moving average are key determinants of transmission | Low temporal and low spatial prediction | Annual | Long history of data is required and a sophisticated analysis that requires a skilled user | SARIMA has great potential to be used as a decision supportive tool due to its ability to predict when and where |
| D19 [32] | Ratio of basic offspring number and basic reproductive ratio is considered outbreak if > = 0.5 | Low temporal and low spatial predictions | N/A | Warrant for more assessment for increasing its sensitivity | Model simulations show that mosquito population are more affected by weather factors than human |
| D20 [33] | Climate factor and incidence rate of dengue before prediction period were superior to rainfall index of week-n | High temporal and low spatial prediction | 2–7 weeks | N/A | The provision of both structure and infrastructure is recommended to be in line with incidence rate prediction value |
| D21 [39] | The model has useful only up to lead 6 times, i.e., correlation >0.5, and as the lead times increase, the match between prediction and observation deteriorates | High temporal but low spatial prediction | 4–24 weeks | Requires long historical data for the evaluation | The model is well suited due to its simplicity in data requirement and computational effort |
| D22 [58] | Predictions are improved both spatially and temporally when using the GLMM; sensitivity of 83% and false alarm of 8% | high temporal and low spatial prediction | 12 weeks | Fails to capture the temporal variability in case counts (due to population immunity to the dominant circulating serotype or specific health interventions) | Seasonal climate forecasts could predict incidence months in advance |
| D23 [37] | Models link outbreaks and climatic conditions and yielded 1-week lag based on spatiotemporal predictions | High temporal and low spatial prediction | 8–12 weeks | N/A | SBME is valuable to timely identify, control, and efficiently prevent disease spreading in time and space |
| D24 [29] | The model was superior (sensitivity: 57%) to the null model (33%) | High temporal and low spatial prediction | 4–12 weeks | N/A | Incorporating real-time seasonal climate forecasts and epidemiological data is beneficial for prediction |
| D25 [69] | The rank probability skill score RPSS was superior to the benchmark with AUC of 0.86 for temperature and 0.84 for rainfall | High temporal and low spatial prediction | 12 weeks | N/A | Close collaboration between public health specialists, climate scientists, and mathematical modelers is crucial for successful implementation of seasonal climate forecasts |
| D26 [36] | The prediction improved with applying criterion >50% chance of exceeding 300 cases per 100,000 inhabitants, with false alarm: 25% | High temporal and low spatial prediction | 12 weeks | N/A | Visualization technique to map ternary probabilistic forecasts can identify areas where the model predicts with certainty a particular disease risk category |
| D27 [56] | The model detected 5 and rejected 14 within 24 months. The Pierce skill score was 0.49, with AUC: 86% and sensitivity: 92% | Low temporal and low spatial prediction | Annual (year-to-year variability) | There is no proper mechanism to track commute-related infections to neighboring districts | Depending upon climatic factors, the previous month’s disease cases had a significant effect on disease incidences of the current month |
| D28 [22] | Sensitivity: 75% at 5 weeks but less sensitive to the outbreak size. Prediction improves when climatic variables and incidence in regions further away from the equator. | High temporal and low spatial prediction | 1–5 weeks | Prediction accuracy might improve if incidence and weather information can be collected at a finer resolution | Short-term LASSO models predictions perform better than longer-term predictions, encouraging public health agencies to respond at short-notice to early warnings |
| Z1 [24] | Integer-valued autoregression is useful predictive model and enhanced by incorporating Google Trends data | High temporal and low spatial prediction | 1–12 weeks | N/A | Accessible and flexible dynamic forecast model can advance early warning prediction |
| Z2 [25] | Average humidity, total rainfall, and maximum temperature were best meteorological factors with prediction lag between 15 and 20 weeks | High temporal and low spatial prediction | 4–20 weeks | The interaction term between the nonlinear smoothing function of time and the spatial function is unavailable in the model, so a real spatiotemporal pattern was unable to be investigated in this study | Meteorological factors are useful for predicting ZIKV epidemics |
| M1 [47] | The model was able to predict both El Niño and non-El Niño malaria outbreaks with high specificity and sensitivity | High temporal prediction | 3–4 months | Data may not be readily available at the district level, and it may not be site specific | Rainfall and unusually high maximum temperatures and the number of inpatient malaria cases 3–4 months later provide a good prediction model |
| M2 [48] | Additive model are most suited for poorly drained U-shaped valley ecosystems while the multiplicative model was most suited for the well-drained V-shaped valley ecosystem | High temporal prediction | 2–4 months | N/A | Additive and multiplicative models are designed for use in the common, well-, and poorly drained valley ecosystems |
| M3 [34] | The models indicated that climate variability remains a major driver of malaria epidemics | High temporal prediction | 2–4 months | N/A | The multiplicative model maintained consistent prediction due to stakeholders’ confidence |
| M4 [49] | Malaria cases exhibited positive associations with LST at a lag of 1 month and positive associations with indicators of moisture at lags between 1 and 3 months | High temporal and low spatial prediction | 1–3 months | Requires weekly rather than monthly intervals, to enable operational forecasting | Integrating modeling approaches based on historical case data (early detection) and environmental data (early warning) can enhance the effectiveness of risk forecasting |
| M5 [50] | Only rainfall had a consistently significant relationship with malaria | High temporal and low spatial prediction | 1–6 months | N/A | Rainfall provides the best predictor of malaria transmission |
| M6 [51] | Temperature is the most relevant climatic parameter thus. Sporogonic and gonotrophic cycles showed to be key entomological variables controlling the transmission | High temporal prediction | 1 month | Too many variables and phases that make it difficult to set in place on daily basis | Environmental factors and climate variability can be merged with selected mathematical tools (statistical and biological/eco-epidemiological models) for improved prediction tool |
| M7 [46] | Within early detection window (the past 6 weeks) and an early warning forecast window (the upcoming 4 weeks), the mean observed or forecasted incidence was classified as being above the mean outbreak threshold, between the mean threshold and the mean expected incidence, or below the mean expected incidence | High temporal and spatial prediction | 1–4 weeks | N/A | Malaria surveillance data and environmental monitoring data can be integrated to enable near real time malaria forecast in the Amhara region |
Reference: Article ID; D = dengue, Z = Zika, and M = malaria.
AUC, area under the curve; CIDARS, China Infectious Disease Automated-alert and Response System; DBSI, Dengue Baidu Search Index; EWS, early warning system; GLMM, generalized linear mixed model; ICC, intraclass correlation coefficient; LASSO, least absolute shrinkage and selection operator; LST, land surface temperature; N/A, not available/no answer; NPV, negative predictive value; PPV, positive predictive value; RMSE, root mean squared error; RPSS, rank probability skill score; SARIMA, seasonal autoregressive integrated moving average; SBME, stochastic Bayesian maximum entropy; SVR, support vector regression; ZIKV, Zika virus.