Binary Screening Model

• Binary screening models are mathematical methods that map input features to a binary outcome using deterministic or probabilistic decision rules.
• They are applied in diverse fields, including high-dimensional variable selection, group testing, online candidate screening, and automated clinical diagnostics.
• These models offer robust error control and theoretical guarantees under extreme sample or dimensionality constraints, ensuring reliable performance.

A binary screening model is a methodological and algorithmic paradigm in which each instance—or item—subject to assessment is classified into one of two possible categories (typically “positive”/“negative” or “included”/“excluded”) based on observed features, covariates, or experimental outcomes. This fundamental structure underlies applications in high-dimensional data analysis, group testing, automated clinical screening, candidate recruitment, systematic review triage, astrophysical event discrimination, and more. Binary screening models are characterized by rigorous decision rules, often under extreme sample, dimensionality, or deployment constraints, and are frequently associated with theoretical guarantees on error rates, consistency, optimality, or robustness.

1. Mathematical Formulation and General Definition

A binary screening model operates by mapping an input vector $x_i$ (which may represent features, attributes, or measurements) to a binary decision $\hat{y}_i \in \{0,1\}$, or equivalently “positive”/“negative”, via a deterministic or probabilistic rule: $\hat{y}_i = f(x_i;\theta)$ where $\theta$ are model parameters, and $f$ is application- and data-specific. The model is trained or specified so as to minimize a loss (such as cross-entropy) or optimize an operational metric (such as AUC, sensitivity, or error rate), possibly under distributional or structural constraints.

Binary screening may be “hard” (deterministic label assignment) or “soft” (probabilistic output compared to a fixed threshold). Paradigms include classical regression screening, group-test combinatorics, online assignment, neural-network classifiers, and linear-programming based decision rules. Training—or tuning—relies on data, task-specific requirements, and explicit constraints to enforce error control or guarantee structure recovery.

2. Applications Across Domains

High-Dimensional Variable Screening

In ultrahigh-dimensional binary regression ($p \gg n$), binary screening is essential for subset selection prior to downstream modeling. Sure Independence Screening (SIS) and its variants compute simple association scores (marginal correlations or MLE slopes from a univariate model) for each feature against the binary response, retaining top-ranked variables for further estimation or inference. Theoretical results ensure that all truly active predictors are retained with high probability under partial orthogonality and sparsity conditions, even when the marginal models are misspecified, provided the predictors are multivariate normal and the true/working link belongs to a normal scale mixture class (Chang, 2014). Recent works extend this to model-free settings using energy distances, providing exact selection under minimal distributional assumptions, and coherent classifiers with provable risk consistency in ultrahigh dimensions (Roy et al., 2022).

Group Testing and Combinatorial Screening

Binary screening models also refer to non-adaptive combinatorial group testing, where the goal is to identify a small subset of “defective” items among a large population via pooled tests. This is formalized as the design of a binary test matrix (code) with the property that the results of $N$ tests permit, with zero or exponentially small error, the identification of all items in the defective set. The key combinatorial object is the $d$-disjunct code, guaranteeing that any up to $d$ positives are perfectly recoverable from test outcomes. The rate $R(d)$ is the limiting ratio of $\log_2 M/N$, and is tightly bounded: $\Omega(1/d^2) \le R(d) \le O((\ln d)/d^2)$, with explicit constructions via Reed-Solomon concatenation and constant-weight codes (D'yachkov, 2014).

Online and Sequential Assignment

Binary screening arises in online matching/assignment tasks, e.g., recruitment or departmental assignment, where each candidate is irrevocably classified (retain/discard) in real time. The aim is to minimize the number of false positives retained while ensuring the retained set contains an optimal matching. With i.i.d. arrivals and no training data, greedy binary screening achieves optimal bounds up to constants: $O(k \log \min\{n/k, k/\delta\})$ for $k$ matched items and confidence $\delta$. With access to a training sample, binary thresholds learned offline allow exponential improvements, down to $O(k[\log d + \log\log(n/k) + \log\log(1/\delta)])$ (Cohen et al., 2019).

Automated Clinical and Biomedical Screening

In clinical screening contexts, such as depression detection from speech, binary screening models ingest raw signal (audio) or text (ASR conversation transcripts) and output a decision indicating the risk category (e.g., PHQ-8 above/below threshold). Robust architectures use deep CNN-LSTM (for acoustics) or LSTM-based LLMs with transfer learning, pooling segment-level predictions to a session-level binary output. Performance is characterized by cross-entropy loss minimization and evaluation by AUC, sensitivity, and specificity. Such models maintain robustness with respect to demographic and session variables, a critical requirement for remote deployment (Lu et al., 2024).

Systematic Review and Document Screening

Binary screening models are employed for large-scale automated inclusion/exclusion of documents in systematic reviews, where LLMs are prompted to make zero-shot “include/exclude” decisions for each abstract, guided by explicit criteria. Metrics include sensitivity, precision, and balanced accuracy, with ensemble approaches (series/parallel rules) yielding perfect recall while managing precision. Review-centric factors (criteria clarity, reporting quality) induce significant performance variation, underscoring the need for domain-specific validation (Sanghera et al., 2024).

Physics: Screening in Binary Astrophysical Systems

In astrophysical binaries under modified gravity with screening mechanisms (e.g., chameleon, Vainshtein, kinetic screening), “binary screening models” describe the modification of orbital dynamics and gravitational waveforms due to spatially varying effective Newton constants. The formalism yields quantitative corrections to the binding energy, power, and phase evolution, allowing stringent constraints on the screening parameter $\mu$ from observed inspiral signals (Honardoost et al., 2019, Renevey et al., 2021, Bezares et al., 2021).

3. Model Architectures and Algorithmic Variants

The structure of a binary screening model depends largely on the scientific context. Example architectural and algorithmic classes include:

• Marginal/Univariate Screening: For each feature $X_k$, compute $|\mathrm{Cov}(Y, X_k)|$ or fit one-dimensional logistic/probit models to define screening priorities (Chang, 2014).
• Energy-based Feature Screening: Marginal or pairwise energy distances ($\mathcal{E}_k$, $\mathcal{E}_{\{i,j\}}$) measure association between distributions across classes, supporting exact screening and subsequent risk-consistent classification (Roy et al., 2022).
• Combinatorial Matrix Codes: The test matrix $X\in\{0,1\}^{N\times t}$ in group testing is engineered for disjunctness, guaranteeing recovery of all $d$-size defective sets (D'yachkov, 2014).
• Online Threshold and Matching Policies: Thresholds-policy retains items passing trained thresholds, sometimes combined with greedy assignment (Cohen et al., 2019).
• Deep Neural Encoders with Transfer Learning: Stacked CNN and LSTM layers pretrained on ASR (acoustic) or generic/domain-adapted text corpora (NLP), followed by session-level aggregation (Lu et al., 2024).
• Ensembles and Human-AI Series/Parallel: Logical combinations (“and”, “or”) of binary decisions from multiple models or human reviewers, controlling trade-offs between sensitivity and precision (Sanghera et al., 2024).
• Linear Programming-based Screening: Reformulating the maximum-score estimation into interval bounding for classification, with abstain/randomize fallback for ambiguity, allowing for minimax optimality and straightforward implementation (Horowitz et al., 25 Jul 2025).

4. Theoretical Guarantees and Statistical Properties

Binary screening models are typically accompanied by rigorous theoretical analyses:

• Sure Screening Property: Probability that all relevant features are retained tends to 1 as $n\to\infty$ under minimal signal conditions, provided the contamination in marginal scores is controlled by predictor dependence structure (Chang, 2014, Roy et al., 2022).
• Error Rate Bounds: In group-testing, error exponents and exact recovery rates are connected to the rate of the underlying superimposed code and can be computed from random coding or explicit combinatorial designs (D'yachkov, 2014).
• Minimax Optimality and Confidence Intervals: In partially identified binary classification, the probability that the model’s classification differs from the oracle is asymptotically bounded above by pre-specified $\alpha$ and a vanishing term as $n$ grows (Horowitz et al., 25 Jul 2025).
• Robustness and Generalization: Empirical results for clinical and document screening tasks show stable performance across user and session subpopulations, with no need for post-hoc tuning or inclusion of explicit metadata (Lu et al., 2024, Sanghera et al., 2024).
• Statistical Inference Post-Screening: Recent advances enable valid hypothesis testing after variable screening, controlling selective type I error in high-dimensional logistic regression (Umezu et al., 2019).

5. Implementation Protocols and Deployment Considerations

Implementation of binary screening models is application-dependent, with deployment protocols typically following these principles:

• Zero-shot and Pretrained Approaches: Modern document and speech screening leverage zero-shot LLMs or deep transfer learning, substantially reducing data annotation requirements (Lu et al., 2024, Sanghera et al., 2024).
• Partitioned Evaluation: Proper evaluation enforces strict separation of training and test sets, often ensuring no overlap of instances (e.g., no speaker overlap), and systematically analyzes performance across demographic, temporal, and operational strata (Lu et al., 2024).
• Logical Aggregations in Ensemble Screening: Model outputs are combined using logical “and”/“or” for trade-offs between recall and precision; nearly all practical deployments rely on domain-specific tuning and validation on held-out data before automation (Sanghera et al., 2024).
• Sequential and Parallel Designs: In large-scale candidate or study screening, deployment can be series (pre-screen with model, then human) or parallel (human and model independently, union or intersection for final inclusion).
• Transparent Reporting and Calibration: For clinical or systematic review impacts, model, prompt, and ensemble details are reported per PRISMA and related evidence synthesis guidelines (Sanghera et al., 2024).
• Computational Practicality: Efficient algorithms (e.g., reformulating NP-hard optimization as simple LPs, $O(pn^2)$ screening for energy distances) are favored for large-scale or ultrahigh-dimensional applications (Horowitz et al., 25 Jul 2025, Roy et al., 2022).

6. Impact, Limitations, and Future Directions

Binary screening models underpin critical infrastructure in high-dimensional inference, large-scale discrimination, and cost-sensitive decision-making. In settings ranging from genomic analysis through remote clinical assessment and systematic review, these models enable scalable, robust, and theoretically well-characterized selection with explicit statistical guarantees.

Current limitations reflect dependence on distributional assumptions or feature dependence (e.g., misspecified links can hinder marginal screening if the contamination is large (Chang, 2014)), computational costs in ultra-large $p$ or combinatorial settings, and the necessity for careful offline/online calibration or domain-specific prompt engineering in automated screening pipelines. Extension to multi-class screening, fully adaptive or “active” designs, and robustification to adversarial distributions or heavy-tailed features remain open areas for research. Expanding selective inference methodologies beyond marginal screens to arbitrary screening rules under weak dependence also remains an open statistical frontier.

Binary screening models provide a unified mathematical and algorithmic foundation for a wide diversity of scientific and operational tasks involving efficient, error-controlled discrimination or selection in two-class regimes, grounded in rigorous statistical, combinatorial, and computational analyses.