Cohort-Squeeze: Cross-Domain Optimization

• Cohort-Squeeze is a cross-disciplinary concept that compresses, reuses, and analyzes cohorts under explicit constraints to enhance performance in diverse fields.
• It is applied in candidate selection, federated learning, mortality analyses, EHR representation, and adaptive trial design using methods like submodular maximization and proximal point techniques.
• Its practical impact is evidenced by improved communication efficiency, transparent auditability, and robust performance guarantees validated through extensive empirical simulations.

Cohort-Squeeze denotes a family of technically distinct but structurally related ideas in which a cohort is compressed, repeatedly reused, or analyzed as a concentrated object under explicit constraints. In cross-device federated learning, it is a communication-efficient training paradigm that reuses the same cohort of clients across multiple consecutive server–client interactions within a global round (Yi et al., 2024). In cohort selection, it denotes compressing a larger, merit-qualified pool into a limited-size final cohort while optimizing diversity in a transparent, auditable way with performance guarantees (Huppenkothen et al., 2019). In mortality analysis, the original papers do not use the phrase, but the supplied interpretation maps it to cohort-specific acceleration or compression of mortality improvements, visible as sharp, high, and narrow peaks in cohort-effect series (Zhang et al., 2015). This suggests that the expression is not a single standardized term, but a recurring way of describing constrained cohort-level optimization, inference, or concentration.

1. Terminological scope

Across the cited literature, the word cohort is stable but the operative meaning of squeeze changes with the problem class. In algorithmic selection, the squeeze is a cardinality-constrained reduction from a qualified pool to a final cohort. In federated learning, it is repeated exploitation of the same sampled cohort before the server rotates to a new one. In mortality, it is a compressed birth-year interval with unusually concentrated cohort effects. In electronic health record analysis, it becomes a squeeze-and-gate operation that compresses cohort context into a patient embedding. In adaptive trial design, it refers to enlarging the next cohort when current evidence indicates proximity to the optimal biologic dose.

A common misconception is to treat these usages as interchangeable. The sources do not support that. Some works use “Cohort-Squeeze” explicitly as a method name or design principle, whereas others only admit the interpretation through supplied technical mappings. This suggests that the phrase functions as a cross-domain abstraction rather than a canonical single-field term.

2. Cohort compression in candidate selection

In candidate selection, Cohort-Squeeze is formalized by Entrofy as the reduction of a merit-qualified set $S$ to a final cohort $X \subset S$ of size $k$, while optimizing pre-declared diversity targets over binary attributes $a_i : S \to \{0,1\}$ with weights $w_i$ and target proportions $p_i$ (Huppenkothen et al., 2019). The procedure is explicitly two-stage. Stage 1 is blind merit review: markers of identity are stripped away, and the committee selects all candidates that meet merit-based criteria. Stage 2 performs diversity optimization on that admissible pool.

The hard formulation,

$\forall i:\ \sum_{x \in X} a_i(x) \ge k \cdot p_i,$

is acknowledged as potentially infeasible. Entrofy therefore uses a soft truncated-coverage objective,

$$f(X)=\sum_i w_i \cdot \min\!\left(k \cdot p_i,\ \sum_{x \in X} a_i(x)\right),$$

and then applies a concave diminishing-returns transformation,

$$f_\alpha(X)=\sum_i w_i \cdot \left[\min\!\left(k \cdot p_i,\ \sum_{x \in X} a_i(x)\right)\right]^\alpha,\qquad 0<\alpha \le 1.$$

The cap implements soft quotas, while $\alpha<1$ favors filling undersupplied attributes first. The experiments recommend $X \subset S$0.

Optimization is performed by greedy maximization under the cardinality constraint $X \subset S$1. At each step, the algorithm computes the marginal gain

$X \subset S$2

for each remaining candidate and adds the maximizer. Exact ties are broken uniformly at random, and randomized near-ties are handled by sampling uniformly from the top $X \subset S$3-quantile by marginal gain. A Monte Carlo multi-start variant repeats this randomized greedy process and returns the cohort with largest objective value.

The key theoretical point is that each truncated coverage term is non-negative, monotone, and submodular, and the summed objective preserves these properties. Under a cardinality constraint, greedy maximization therefore attains a $X \subset S$4 approximation to the optimal value, while exact optimization is NP-Hard. Runtime is $X \subset S$5 for $X \subset S$6 and $X \subset S$7 binary attributes after encoding.

The empirical program is unusually extensive. The simulations comprise 24.3 million runs total. Single-run failure is most likely when an attribute appears only in the planted optimal solution and many distractors are added, with a phase transition in failure rates around at least 100 distractors. In a difficult case with $X \subset S$8, 105 candidates, targets $X \subset S$9, and distractors with $k$0 for one attribute, single-run failure is about 49%, but multi-start Entrofy drives failure to approximately 0 by $k$1. In the Astro Hack Week case study, overall distance to targets improved from $k$2 to $k$3 in 2016 and from $k$4 to $k$5 in 2017; post-workshop surveys indicated strong agreement that algorithmic selection is more transparent and fair. The intended significance is not replacement of human merit judgment, but separation of merit assessment from diversity optimization in a way that is transparent and auditable.

3. Multi-round cohort reuse in cross-device federated learning

In cross-device federated learning, Cohort-Squeeze challenges the standard design primitive in which a sampled cohort participates in a single communication round and is then discarded (Yi et al., 2024). The paper’s central claim is that one can “squeeze more juice” out of each cohort by reusing the same sampled clients across multiple consecutive server–client interactions within a global round.

The baseline optimization problem is

$k$6

Instead of a single FedAvg-style interaction, the method uses a federated instantiation of a stochastic proximal point method with arbitrary sampling, SPPM-AS. For a sampled cohort $k$7, the exact update is

$k$8

where

$k$9

under importance weights induced by the sampling scheme. Within a global iteration, the same cohort performs $a_i : S \to \{0,1\}$0 sub-rounds to approximately solve the proximal subproblem. This is the core FL meaning of Cohort-Squeeze.

The framework supports several sampling procedures. NICE sampling draws a minibatch uniformly without replacement. Block sampling draws an entire client block. Stratified sampling draws one client from each cluster and is analytically important because it reduces the variance parameter

$a_i : S \to \{0,1\}$1

The theory assumes differentiability and strong convexity of each $a_i : S \to \{0,1\}$2, but does not assume $a_i : S \to \{0,1\}$3-smoothness. With arbitrary $a_i : S \to \{0,1\}$4, the method satisfies

$a_i : S \to \{0,1\}$5

The first term is a linear convergence factor, while the second is the convergence neighborhood. Larger $a_i : S \to \{0,1\}$6 improves the rate but enlarges that neighborhood. Under interpolation, where $a_i : S \to \{0,1\}$7, arbitrarily large $a_i : S \to \{0,1\}$8 gives arbitrarily fast convergence to $a_i : S \to \{0,1\}$9.

The practical motivation is communication efficiency. In the standard setting, the reported reduction in total communication cost needed to reach a target accuracy is up to 74.36%. In hierarchical FL, where local communications are cheaper than global communications, the reported reduction is up to 94.87%. The effect is strongest when local sub-rounds are inexpensive relative to global synchronization and when stratified sampling is used to reduce cohort-induced variance. The main limitation is system-level rather than analytical: reusing a cohort for multiple sub-rounds requires devices to remain available longer within a global iteration, so straggling and dropout become more consequential.

4. Concentrated cohort effects in mortality analysis

In mortality research, Cohort-Squeeze refers to cohort-specific compression in mortality improvements. The differential-geometric formulation represents mortality as a smooth surface $w_i$0 defined by $w_i$1, where $w_i$2 is calendar time, $w_i$3 is age, and $w_i$4 is mortality (Zhang et al., 2015). For a fixed birth year, the cohort curve follows points satisfying $w_i$5 on the discrete age–period grid. The central quantity is the cohort effect index,

$w_i$6

and in the discrete setting,

$w_i$7

A squeeze appears as a sharp peak or narrow high plateau in the $w_i$8 series, together with shorter generation gaps.

The empirical results identify strong cross-country variation. For the U.K., the cohort-effect series shows “two obvious peaks,” with one about 3–5 years and another about 10 years, and the generation born between 1930 and 1935 appears as an approximately plateau-like effect. The reported coefficient of variation for the restricted CEI series, called AICE, is 0.617719261 for the U.K., 0.578148277 for Canada, 0.446035861 for the U.S., and 0.431667221 for Japan. The interpretation supplied with the paper is that larger AICE indicates stronger concentration of cohort effects.

A second line of work embeds cohort effects directly in stochastic mortality models. In the Bayesian state-space formulation,

$w_i$9

with $p_i$0 following a random walk with drift and the cohort factor $p_i$1 embedded through a shift structure and an AR(1) driver (Fung et al., 2017). Cohort squeeze is diagnosed through sharp deviations in $p_i$2, diagonal residual bands under period-only models, and materially better fit for cohort models. For example, the reported conditional DIC for U.K. males is $p_i$3 for Lee–Carter, $p_i$4 for the simplified cohort model, and $p_i$5 for the full cohort model.

A related parsimonious compromise is the Lee–Carter + Cohort model,

$p_i$6

which fixes the cohort loading to unity rather than fitting an age-specific multiplier (Debonneuil, 2010). The stated purpose is to preserve age-adjusted period improvements while avoiding the overfitting and identifiability problems seen in richer cohort models. In the supplied synthesis, this is presented as a mitigation of the cohort-squeeze tension between age, period, and cohort components.

Taken together, these approaches treat cohort squeeze as a measurable diagonal structure in mortality surfaces, not merely as a descriptive demographic phrase. The controversy is not whether such structure exists, but how it should be regularized, identified, and projected.

5. Cohort intelligence and squeeze-and-gate representation learning in EHR analysis

In electronic health record analysis, the relevant framework is CORE, a universal COhort Representation lEarning architecture that augments patient embeddings with fine-grained cohort information (Liu et al., 2023). Patient-level features are first learned through a diagnosis-informed pre-context task. Diagnoses use Node2Vec embeddings on ontology graphs and GloVe embeddings on diagnosis-code semantics; visit features are aggregated and then passed through reverse time attention to form patient embeddings $p_i$7. Agglomerative clustering over $p_i$8 yields cohort assignments.

Within this framework, the supplied material defines a concrete Cohort-Squeeze module. For patient $p_i$9, intra-cohort context is formed from neighbors in the same cohort using cosine similarity and softmax weights,

$\forall i:\ \sum_{x \in X} a_i(x) \ge k \cdot p_i,$0

Optionally, inter-cohort context is obtained from a cohort-centroid graph encoded by a GCN. The squeeze stage computes

$\forall i:\ \sum_{x \in X} a_i(x) \ge k \cdot p_i,$1

followed by a gate

$\forall i:\ \sum_{x \in X} a_i(x) \ge k \cdot p_i,$2

and then updates the patient representation by

$\forall i:\ \sum_{x \in X} a_i(x) \ge k \cdot p_i,$3

or by adding a projected cohort context term. This is explicitly described as a lightweight alternative to CORE’s scalar attention fusion.

The empirical backdrop is strong. On MIMIC-III, ClinicalBERT improves from AUPRC 0.630 to 0.658 under CORE, Med2Vec from 0.540 to 0.605, and MiME from 0.542 to 0.595. On UCI Diabetes, Med2Vec improves from 0.533 to 0.552 and MiME from 0.531 to 0.555. Ablation shows that removing the pre-context module lowers AUPRC and accuracy by approximately 2.3% on MIMIC-III. The broader significance is that the cohort is not merely a sampling device; it becomes an explicit latent structure used to denoise and recode individual representations.

6. Adaptive cohort sizing in BOIN12 phase I/II design

In Phase I/II oncology design, Cohort-Squeeze is an adaptive cohort size determination method layered on BOIN12 to shorten trial duration by expanding the next cohort when the upcoming dose appears close to the optimal biologic dose (Kojima, 2023). BOIN12 combines toxicity and efficacy using utilities: 100 for no toxicity and efficacy, 60 for toxicity and efficacy, 40 for no toxicity and no efficacy, and 0 for toxicity and no efficacy. The desirability value is

$\forall i:\ \sum_{x \in X} a_i(x) \ge k \cdot p_i,$4

The expansion trigger is the desirability probability

$\forall i:\ \sum_{x \in X} a_i(x) \ge k \cdot p_i,$5

The operating thresholds are explicit. The maximum acceptable toxicity probability is $\forall i:\ \sum_{x \in X} a_i(x) \ge k \cdot p_i,$6, the minimum acceptable efficacy probability is $\forall i:\ \sum_{x \in X} a_i(x) \ge k \cdot p_i,$7, the safety admissibility cutoff is $\forall i:\ \sum_{x \in X} a_i(x) \ge k \cdot p_i,$8, and the efficacy admissibility cutoff is $\forall i:\ \sum_{x \in X} a_i(x) \ge k \cdot p_i,$9. The recommended expansion threshold is $$f(X)=\sum_i w_i \cdot \min\!\left(k \cdot p_i,\ \sum_{x \in X} a_i(x)\right),$$0, with $$f(X)=\sum_i w_i \cdot \min\!\left(k \cdot p_i,\ \sum_{x \in X} a_i(x)\right),$$1 as a stricter alternative. If $$f(X)=\sum_i w_i \cdot \min\!\left(k \cdot p_i,\ \sum_{x \in X} a_i(x)\right),$$2 and the observed counts satisfy the precomputed expansion table, the next cohort size is increased from 3 to 6; otherwise it remains 3. Expansion is blocked if it would exceed the maximum trial sample size $$f(X)=\sum_i w_i \cdot \min\!\left(k \cdot p_i,\ \sum_{x \in X} a_i(x)\right),$$3.

The simulation design uses 5 dose levels, $$f(X)=\sum_i w_i \cdot \min\!\left(k \cdot p_i,\ \sum_{x \in X} a_i(x)\right),$$4, accrual of 3 patients per month, a toxicity window of 45 days, an efficacy window of 60 days, 16 scenarios, and 10,000 simulated trials per scenario. With the stopping rule, adaptive BOIN12 reduces duration by 3.2 months on average, a mean 14.2% reduction relative to fixed-size BOIN12. Without the stopping rule, the reduction is 5.7 months on average, a mean 19.0% reduction. The mean difference in correct optimal biologic dose selection is small: $$f(X)=\sum_i w_i \cdot \min\!\left(k \cdot p_i,\ \sum_{x \in X} a_i(x)\right),$$5 in Case A and $$f(X)=\sum_i w_i \cdot \min\!\left(k \cdot p_i,\ \sum_{x \in X} a_i(x)\right),$$6 in Case B. In the paper’s worked examples with $$f(X)=\sum_i w_i \cdot \min\!\left(k \cdot p_i,\ \sum_{x \in X} a_i(x)\right),$$7, expanding two cohorts to size 6 reduces trial time by 100 days in one example and 50 days in another. Here the squeeze is calendar-time compression through adaptive batch enlargement, not statistical shrinkage.

7. Additional domain-specific meanings and cross-cutting themes

Two additional uses further illustrate the breadth of the term. In Chesapeake Bay oyster reef analysis, cohort squeeze refers to age-truncation: older age classes are rare or absent and most individuals are concentrated in the youngest classes (Griffin et al., 8 Dec 2025). The study fits reef-year Gaussian mixture models to shell-length distributions and links age groups across time under a non-shrinkage rule,

$$f(X)=\sum_i w_i \cdot \min\!\left(k \cdot p_i,\ \sum_{x \in X} a_i(x)\right),$$8

The reported ecological finding is that age-truncation eased in the mid-to-late 2010s, with cohorts more frequently reaching ages 4–5 and some age-6 classes identified in the James and Piankatank rivers.

In large case-cohort survival studies, the phrase is used for computational compression. Influence function-based supersampling selects a small, informative supersample for multiple imputation, using

$$f(X)=\sum_i w_i \cdot \min\!\left(k \cdot p_i,\ \sum_{x \in X} a_i(x)\right),$$9

as inclusion probabilities, balanced sampling, and calibrated weights (Kim et al., 18 Nov 2025). The stated result is efficiency comparable to imputing the entire cohort even with a small supersample, together with runtime reductions of about 95% in simulations and about 96% in the NIH–AARP application.

These additional usages reinforce a common structure. A cohort is treated as a higher-level object whose size, reuse, concentration, or representational role is manipulated under explicit constraints. This suggests a cross-disciplinary abstraction: Cohort-Squeeze names operations that extract more utility, information, or interpretive resolution from cohort structure than a baseline one-shot procedure would permit. The exact mathematics differ—submodular maximization, proximal point methods, curvature-based cohort diagnostics, GCN-based recoding, Bayesian desirability rules, Gaussian mixtures, or influence-function-based supersampling—but the recurrent theme is the same: cohort structure is not incidental metadata; it is the primary locus of optimization or inference.