Artificial Intelligence Foundations for Public Health
Expert-defined terms from the Certificate Programme in AI for Public Health Surveillance course at London School of Planning and Management. Free to read, free to share, paired with a professional course.
Artificial Intelligence (AI) – A broad field encompassing computational t… #
Related terms: Machine Learning, Deep Learning, Natural Language Processing. AI systems process large volumes of health data to identify patterns, predict disease spread, and support decision‑making. Example: An AI model predicts influenza peaks using historical case counts and weather data. Challenges: Data quality, model transparency, and ethical considerations such as privacy and bias.
Algorithmic Bias – Systematic errors that cause an AI model to produce pr… #
Related terms: Fairness, Discrimination, Data Imbalance. In public health, biased algorithms may under‑detect outbreaks in underserved communities. Example: A predictive model trained on urban hospital data fails to flag rural disease clusters. Challenges: Identifying hidden biases, correcting imbalanced datasets, and ensuring equitable health interventions.
Anomaly Detection – Techniques that identify data points deviating marked… #
Related terms: Outlier Detection, Change Point Detection, Surveillance Alerts. Used in syndromic surveillance to flag unusual spikes in symptom reports. Example: A sudden rise in emergency department visits for respiratory distress triggers an alert. Challenges: Distinguishing true public health events from noise, handling seasonal variations, and setting appropriate thresholds.
Automated Text Mining – The use of algorithms to extract structured infor… #
Related terms: Natural Language Processing, Entity Recognition, Sentiment Analysis. Enables rapid identification of emerging health threats from web streams. Example: Mining Twitter for mentions of “fever” and “cough” to map a possible COVID‑19 flare‑up. Challenges: Language variability, sarcasm detection, and ensuring data privacy.
Bayesian Inference – A statistical framework that updates the probability… #
Related terms: Prior Distribution, Posterior Probability, Markov Chain Monte Carlo. In disease modeling, Bayesian methods incorporate prior knowledge about transmission rates. Example: Updating the estimated reproduction number (R0) as daily case counts arrive. Challenges: Choosing appropriate priors, computational intensity, and communicating uncertainty to policymakers.
Bias‑Variance Trade‑off – The balance between a model’s ability to captur… #
Related terms: Overfitting, Underfitting, Model Complexity. Selecting the right regularization technique is crucial for reliable public‑health predictions. Example: A highly complex neural network fits past outbreak data perfectly but performs poorly on future seasons. Challenges: Diagnosing the source of error and adjusting model capacity accordingly.
Case‑Based Reasoning – An AI approach that solves new problems by adaptin… #
Related terms: Nearest Neighbor, Instance‑Based Learning, Knowledge Retrieval. Useful for rapid response when a novel disease resembles a known pathogen. Example: Comparing a new respiratory syndrome to past SARS cases to infer likely transmission routes. Challenges: Maintaining a comprehensive case repository and ensuring relevance of past cases.
Convolutional Neural Network (CNN) – A deep learning architecture special… #
Related terms: Filters, Pooling, Feature Maps. In public health, CNNs analyze radiological images to detect tuberculosis or COVID‑19 pneumonia. Example: A CNN classifies chest X‑rays with >90% accuracy, supporting triage in low‑resource settings. Challenges: Need for large annotated datasets, interpretability of learned features, and computational resources.
Cross‑Validation – A resampling method for evaluating model performance b… #
Related terms: K‑Fold, Hold‑Out, Model Assessment. Ensures that predictive models for disease incidence are robust across different temporal or geographic splits. Example: Five‑fold cross‑validation of a random‑forest model predicting dengue cases across provinces. Challenges: Maintaining temporal integrity (e.G., Not training on future data) and handling imbalanced outcome classes.
Data Governance – Policies, procedures, and standards that ensure data is… #
Related terms: Data Stewardship, Privacy, Compliance. Critical for public‑health surveillance systems that aggregate personal health records. Example: Implementing role‑based access controls for a national disease‑reporting platform. Challenges: Balancing data availability for research with legal and ethical constraints, especially under differing jurisdictional regulations.
Data Imbalance – Situations where the number of observations in one class… #
Related terms: Oversampling, Undersampling, Synthetic Minority Over‑sampling Technique (SMOTE). Imbalance can cause predictive models to ignore minority events. Example: Using SMOTE to augment training data for Ebola case detection. Challenges: Avoiding over‑fitting synthetic samples and preserving the true distribution of rare events.
Data Integration – The process of combining data from disparate sources i… #
Related terms: Data Fusion, ETL (Extract‑Transform‑Load), Interoperability. Enables comprehensive surveillance by linking laboratory reports, electronic health records, and mobility data. Example: Merging hospital admission logs with sewage‑monitoring results to improve early detection of viral outbreaks. Challenges: Heterogeneous formats, differing terminologies, and real‑time synchronization.
Decision Support System (DSS) – Computer‑based tools that assist public‑h… #
Related terms: Dashboard, Alert System, Scenario Modeling. AI‑enhanced DSS can simulate the impact of vaccination strategies. Example: A DSS suggests optimal allocation of limited COVID‑19 vaccine doses across districts based on projected infection curves. Challenges: Ensuring usability, avoiding information overload, and maintaining up‑to‑date models.
Deep Learning – A subset of machine learning that employs multi‑layered n… #
Related terms: Neural Networks, Representation Learning, Backpropagation. In public health, deep models process high‑dimensional data such as genomic sequences or satellite imagery. Example: A deep recurrent network predicts malaria incidence using climate, land‑use, and travel patterns. Challenges: Large computational demands, lack of transparency, and requirement for extensive labeled datasets.
Dimensionality Reduction – Techniques that compress high‑dimensional data… #
Related terms: Principal Component Analysis (PCA), t‑Distributed Stochastic Neighbor Embedding (t‑SNE), Autoencoders. Facilitates visualization and reduces overfitting in models using many health indicators. Example: Applying PCA to reduce 200 environmental variables to 10 principal components before feeding them into a predictive model. Challenges: Interpreting reduced dimensions and selecting the appropriate number of components.
Ensemble Learning – Combining predictions from multiple models to improve… #
Related terms: Bagging, Boosting, Stacking. Frequently used in outbreak forecasting to capture diverse patterns. Example: An ensemble of random forests, gradient‑boosted trees, and support vector machines predicts weekly influenza cases with higher accuracy than any single model. Challenges: Increased computational cost, risk of correlated errors, and difficulty in interpreting the final ensemble decision.
Explainable AI (XAI) – Methods that make the inner workings of AI models… #
Related terms: Interpretability, Model Transparency, Feature Importance. In public health, XAI builds trust among clinicians and policymakers. Example: Using SHAP values to show that air‑quality indices contributed most to a predicted rise in asthma exacerbations. Challenges: Balancing explanation depth with model performance and presenting insights in a non‑technical manner.
Feature Engineering – The process of creating, selecting, and transformin… #
Related terms: Variable Construction, Encoding, Scaling. In disease surveillance, engineered features may include lagged case counts, mobility indices, or vaccination coverage rates. Example: Adding a 7‑day moving average of reported symptoms as a predictor for next‑day outbreak risk. Challenges: Avoiding leakage of future information, ensuring reproducibility, and handling high‑frequency time‑series data.
Geospatial Analysis – Techniques that incorporate geographic location int… #
Related terms: GIS (Geographic Information Systems), Spatial Autocorrelation, Heatmaps. Enables identification of disease hotspots and resource allocation. Example: Mapping dengue cases to pinpoint neighborhoods with vector‑breeding sites for targeted insecticide spraying. Challenges: Spatial data privacy, varying resolution of geographic layers, and accounting for population density differences.
Gradient Boosting Machine (GBM) – An ensemble learning method that builds… #
Related terms: XGBoost, LightGBM, AdaBoost. GBMs often achieve state‑of‑the‑art performance in predictive public‑health tasks. Example: Using XGBoost to forecast weekly COVID‑19 hospital admissions based on case counts, vaccination rates, and mobility metrics. Challenges: Hyper‑parameter tuning, risk of overfitting, and computational intensity on large datasets.
Health Informatics – The interdisciplinary study of the design, developme… #
Related terms: Electronic Health Records (EHR), Clinical Decision Support, Interoperability. Forms the backbone for data collection in AI‑driven surveillance. Example: Integrating EHR data streams with a national reporting system to enable near‑real‑time influenza monitoring. Challenges: Standardizing terminologies (e.G., SNOMED, ICD), ensuring data security, and addressing legacy system constraints.
Hybrid Modeling – Combining mechanistic (theory‑driven) models with data‑… #
Related terms: Compartmental Models, Machine Learning, Model Fusion. Improves forecasting when epidemiological theory alone is insufficient. Example: Embedding a SEIR (Susceptible‑Exposed‑Infectious‑Recovered) structure within a neural network that learns residual patterns from real‑world case data. Challenges: Aligning model scales, calibrating parameters jointly, and preventing contradictions between components.
Imbalanced Classification – A specific case of data imbalance where the t… #
Related terms: Cost‑Sensitive Learning, Focal Loss, ROC‑AUC. Adjusting loss functions or sampling strategies can improve detection rates. Example: Applying focal loss to a CNN that identifies images of disease‑specific rash, increasing sensitivity for the minority class. Challenges: Selecting appropriate evaluation metrics and avoiding inflated performance due to trivial majority‑class predictions.
Incident Reporting System – Digital platforms that capture and transmit i… #
Related terms: Surveillance System, Notification, Case Definition. AI can automate triage and prioritize alerts. Example: An AI module classifies incoming reports into “high”, “medium”, or “low” priority based on symptom severity and geographic spread. Challenges: Ensuring completeness of reports, dealing with inconsistent terminology, and preventing false alarms.
K‑Nearest Neighbors (KNN) – A simple, instance‑based learning algorithm t… #
Related terms: Distance Metric, Lazy Learning, Instance Retrieval. Useful for quick baseline models in outbreak detection. Example: Classifying a region’s risk level by comparing recent case trajectories to those of similar regions. Challenges: Sensitivity to noise, computational cost for large datasets, and choosing an appropriate *k*.
Knowledge Graph – A network‑based representation of entities (e #
G., Diseases, symptoms, locations) and their relationships. Related terms: Ontology, Triple Store, Semantic Reasoning. Enables AI to reason over complex public‑health concepts. Example: A knowledge graph links COVID‑19 variants to vaccine efficacy data, supporting policy decisions on booster recommendations. Challenges: Curating accurate relationships, scaling graph queries, and integrating heterogeneous data sources.
Logistic Regression – A classic statistical method for modeling binary ou… #
Related terms: Generalized Linear Model, Maximum Likelihood, Regularization. Frequently used as a baseline for disease risk prediction. Example: Predicting the probability of a patient developing severe COVID‑19 based on age, comorbidities, and vaccination status. Challenges: Handling non‑linearity, multicollinearity, and limited predictive power when relationships are complex.
Machine Learning (ML) – Algorithms that enable computers to learn pattern… #
Related terms: Supervised Learning, Unsupervised Learning, Reinforcement Learning. In public health, ML powers outbreak forecasting, risk stratification, and resource optimization. Example: A random‑forest model predicts weekly dengue incidence using climate, vector‑control, and socioeconomic variables. Challenges: Data sparsity, model interpretability, and maintaining performance across changing epidemiological contexts.
Model Drift – The degradation of a model’s predictive performance over ti… #
Related terms: Concept Drift, Temporal Validation, Retraining. Surveillance models must be monitored continuously to detect drift. Example: A COVID‑19 case‑forecasting model trained on pre‑vaccination data underestimates cases after mass immunization campaigns. Challenges: Establishing drift detection thresholds, scheduling timely model updates, and preserving historical comparability.
Natural Language Processing (NLP) – A suite of methods for analyzing and… #
Related terms: Tokenization, Named Entity Recognition, Sentiment Analysis. Enables extraction of health indicators from free‑text sources such as clinical notes or social media. Example: An NLP pipeline identifies mentions of “loss of smell” in electronic health records to flag possible COVID‑19 cases. Challenges: Dealing with medical jargon, multilingual corpora, and privacy‑preserving text handling.
Neural Network – A computational model composed of interconnected nodes (… #
Related terms: Activation Function, Backpropagation, Deep Architecture. Forms the basis of most modern AI tools in public health. Example: A feed‑forward network predicts hospital bed occupancy from daily admission counts and community transmission rates. Challenges: Selecting appropriate architecture, preventing overfitting, and ensuring computational efficiency.
Outbreak Detection Algorithm – Statistical or AI‑based methods designed t… #
Related terms: CUSUM (Cumulative Sum), EWMA (Exponentially Weighted Moving Average), Syndromic Surveillance. Example: A CUSUM algorithm flags a significant rise in respiratory‑illness reports from a city’s emergency departments. Challenges: Balancing sensitivity and specificity, handling reporting delays, and integrating multiple data streams.
Precision Public Health – The application of high‑resolution data and ana… #
Related terms: Targeted Intervention, Granular Data, Micro‑Stratification. AI enables identification of hyper‑local risk factors. Example: Using machine‑learning models to allocate mosquito‑control resources to neighborhoods with the highest predicted vector density. Challenges: Data privacy at fine geographic scales, ensuring equitable access, and avoiding stigmatization.
Probabilistic Graphical Model – A framework that represents random variab… #
Related terms: Bayesian Network, Markov Random Field, Inference. Useful for modeling complex causal pathways in disease transmission. Example: A Bayesian network incorporates vaccination coverage, contact patterns, and pathogen virulence to estimate outbreak probability. Challenges: Learning accurate structure from limited data and performing efficient inference on large networks.
Public Health Surveillance System – Organized collection, analysis, and d… #
Related terms: Indicator-Based Surveillance, Event‑Based Surveillance, Reporting Network. AI augments these systems with automated anomaly detection and predictive analytics. Example: An AI‑driven platform ingests laboratory test results, hospital admissions, and pharmacy sales to generate weekly situational reports. Challenges: Ensuring data timeliness, standardizing case definitions, and maintaining system sustainability.
Random Forest – An ensemble learning method that constructs multiple deci… #
Related terms: Bagging, Feature Importance, Out‑of‑Bag Error. Frequently employed for classification of disease status and variable selection. Example: A random‑forest model identifies key climatic predictors for malaria incidence across multiple regions. Challenges: Interpreting large numbers of trees, handling correlated features, and tuning the number of trees for optimal performance.
Recurrent Neural Network (RNN) – A class of neural networks designed for… #
Related terms: LSTM (Long Short‑Term Memory), GRU (Gated Recurrent Unit), Time‑Series Forecasting. RNNs excel at modeling disease trajectories over time. Example: An LSTM network predicts the next 14 days of COVID‑19 case counts using past daily reports and mobility metrics. Challenges: Vanishing gradients, requirement for large sequential datasets, and high computational cost.
Reinforcement Learning (RL) – A learning paradigm where an agent interact… #
Related terms: Policy, Q‑Learning, Exploration‑Exploitation. In public health, RL can optimize allocation of limited resources such as vaccines or testing kits. Example: An RL agent learns a vaccination schedule that minimizes projected infection peaks while respecting supply constraints. Challenges: Defining appropriate reward structures, ensuring ethical decision‑making, and dealing with delayed feedback.
Risk Stratification – The process of categorizing individuals or populati… #
Related terms: Scoring System, Predictive Model, Triage. AI enhances stratification by integrating diverse data sources. Example: A gradient‑boosted model assigns high, medium, or low risk of severe influenza to patients based on age, comorbidities, and recent exposure history. Challenges: Avoiding discrimination, updating risk scores as epidemiology evolves, and communicating risk to end‑users.
Scalable Architecture – System designs that can handle increasing data vo… #
Related terms: Distributed Computing, Cloud Infrastructure, Parallel Processing. Essential for national surveillance platforms ingesting millions of records daily. Example: Deploying a Spark‑based pipeline that processes streaming hospital admission data in near real‑time for outbreak alerts. Challenges: Managing cost, ensuring data security across distributed nodes, and maintaining fault tolerance.
Semantic Interoperability – The ability of different information systems… #
Related terms: Ontology, Standardized Vocabulary, HL7 FHIR. Critical for integrating heterogeneous public‑health datasets. Example: Mapping local disease codes to the WHO ICD‑10 ontology enables cross‑border outbreak comparison. Challenges: Aligning multiple terminologies, handling legacy data, and achieving consensus among stakeholders.
Sentinel Surveillance – Focused monitoring of selected health facilities… #
Related terms: Sentinel Sites, Representative Sampling, Early Warning. AI can automate analysis of sentinel data streams. Example: An AI model evaluates weekly influenza‑like illness reports from sentinel clinics to forecast national epidemic curves. Challenges: Ensuring sentinel sites remain representative, dealing with limited sample size, and integrating sentinel data with other surveillance streams.
Supervised Learning – A machine‑learning paradigm where models are traine… #
Related terms: Labels, Training Set, Regression, Classification. Most predictive public‑health models, such as disease incidence forecasts, rely on supervised techniques. Example: Training a support‑vector machine to classify regions as “outbreak” or “non‑outbreak” based on recent case trends and environmental factors. Challenges: Obtaining high‑quality labeled data, preventing overfitting, and adapting models to shifting patterns.
Time‑Series Forecasting – Methods that predict future values based on pre… #
Related terms: ARIMA, Prophet, Seasonal Decomposition. Central to anticipating disease peaks and resource needs. Example: An ARIMA model projects weekly measles cases using five years of historical surveillance data. Challenges: Capturing non‑linear trends, accounting for interventions (e.G., Vaccination campaigns), and handling missing observations.
Transfer Learning – Leveraging knowledge acquired from one task or domain… #
Related terms: Pre‑training, Fine‑tuning, Domain Adaptation. Enables rapid development of models for emerging diseases by reusing weights from existing models. Example: Fine‑tuning a CNN trained on chest X‑rays for pneumonia to detect COVID‑19‑related abnormalities with limited new images. Challenges: Risk of negative transfer when source and target domains differ substantially, and deciding which layers to freeze.
Uncertainty Quantification – Techniques that assess the confidence or var… #
Related terms: Confidence Intervals, Predictive Distribution, Bayesian Modeling. Provides decision‑makers with risk‑aware information. Example: Reporting a 95% credible interval for projected dengue cases, allowing planners to prepare for worst‑case scenarios. Challenges: Computational overhead, communicating uncertainty to non‑technical audiences, and integrating uncertainty into policy thresholds.
Unsupervised Learning – Algorithms that infer structure from data without… #
Related terms: Clustering, Anomaly Detection, Latent Dirichlet Allocation. In surveillance, unsupervised methods can discover novel disease patterns. Example: Applying K‑means clustering to symptom reports to reveal previously unrecognized syndromes. Challenges: Determining the number of clusters, interpreting results, and validating findings against clinical reality.
Validation Cohort – An independent dataset used to assess the generalizab… #
Related terms: External Validation, Hold‑out Set, Generalization Error. Ensures that AI tools perform reliably across different populations. Example: Testing a COVID‑19 severity model developed in Country A on patient data from Country B. Challenges: Accessing comparable external data, handling differences in data collection practices, and avoiding over‑optimistic performance estimates.
Vector‑Borne Disease Modeling – Computational approaches that simulate th… #
G., Mosquitoes). Related terms: Entomological Index, Climate Suitability, Agent‑Based Model. AI enhances these models by incorporating high‑resolution environmental data. Example: A random‑forest model predicts weekly Aedes aegypti abundance using temperature, precipitation, and land‑cover variables, feeding into a dengue transmission simulation. Challenges: Limited vector surveillance data, complex life‑cycle dynamics, and uncertainty in climate forecasts.
Web Scraping – Automated extraction of information from websites, often u… #
Related terms: Crawling, Data Harvesting, API Integration. Provides additional sources for early outbreak signals. Example: Scraping online pharmacy listings for increased sales of antipyretics as a proxy for fever prevalence. Challenges: Legal restrictions, website structure changes, and ensuring data quality.
Zero‑Shot Learning – A paradigm where a model can recognize classes it ha… #
Related terms: Few‑Shot Learning, Attribute Embedding, Transfer Learning. In public health, useful for detecting novel pathogens with limited data. Example: A model trained on known viral image patterns predicts the presence of a newly emerging virus based on its genetic similarity to known families. Challenges: Obtaining reliable auxiliary information, ensuring robustness, and preventing misclassification of truly unknown entities.