Computational Phenotyping of Gender

• Computational phenotyping of gender is the use of algorithmic methods to infer and measure gender-related traits from proxy data sources.
• It spans diverse modalities such as names, images, language, and health records, employing models like CNNs, probabilistic estimators, and latent-variable approaches.
• Research in this area highlights challenges like bias, temporal shifts, cultural variation, and ethical concerns in interpreting gender data.

Computational phenotyping of gender is the use of algorithmic methods to infer, quantify, or audit gender-related structure from indirect data rather than from an explicitly recorded, self-reported gender field. Across recent work, the object of inference ranges from individual labels derived from names, faces, iris or periocular imagery, dialogue, behavioral traces, and electronic health records, to population-level gender composition estimated from name distributions, to system-level audits of how LLMs themselves construct gender (Hafner et al., 20 May 2025). The term therefore denotes not one task but a family of measurement problems whose operational targets vary sharply: some studies use binary male/female or man/woman labels, some add a non-binary class, some estimate group representativeness or disparity, and some treat gender as a latent ontology encoded in a model rather than as a ground-truth attribute of a person (Gronsbell et al., 19 Aug 2025).

1. Conceptual scope and operational targets

The central methodological fact is that computational phenotyping of gender does not have a single stable definition. In name-based work, gender is treated as a proxy category inferred from culturally learned associations between names and gendered group membership, typically because direct gender data are unavailable at scale (Buskirk et al., 2022). In EHR research, phenotypes are built from diagnosis codes, medications, procedures, structured fields, and clinical notes, but the resulting algorithm may stand in for gender identity, gender expression, medical transition, or some mixture of them (Gronsbell et al., 19 Aug 2025). In language-model auditing, the target is neither a person nor a population; rather, it is the model’s implicit representation of gender, including whether it distinguishes gender from sex, meaningfully represents gender-diverse terms, and avoids pathologizing associations (Hafner et al., 20 May 2025).

Operationalizations also differ at the label level. Many studies instantiate a binary classification problem, such as predicting $y \in \{\text{male}, \text{female}\}$ from iris images, names, or facial regions (Kuehlkamp et al., 2017). Other work uses man and woman because the data source is self-report and birth sex or cis/trans status cannot be inferred reliably (Lokala et al., 2022). A third line of work extends image classification from a binary label set to a three-class system by adding Non-binary as a class (Wu et al., 2020). By contrast, the audit of pretrained LLMs probes a set of seven gender identifiers—$$\{a\ man, a\ woman, transgender, nonbinary, genderqueer, genderfluid, two\text{-}spirit\}$$—and evaluates whether sexed contexts induce binary completions and whether gender-diverse terms are erased or pathologized (Hafner et al., 20 May 2025).

A further distinction concerns the unit of analysis. Some studies phenotype individuals from names, text, or images. Others phenotype aggregates. The global gender estimation method estimates the gender composition of a whole list of names as a population-level latent variable rather than inferring each person independently (Antonoyiannakis et al., 2023). The Sarafina score measures gender asset disparity in land acquisition and land tenure security as a dynamic, policy-adjusted quantity, not as an individual attribute (Ogundare et al., 2024). In media analysis, gender is operationalized as visible on-screen presence through the female face ratio (ffr) rather than identity or speech (Mazieres et al., 2020). This suggests that “computational phenotyping of gender” is best understood as an umbrella for heterogeneous inferential regimes linked by their reliance on proxy variables and algorithmic measurement.

2. Modalities, pipelines, and representative model classes

The methodological diversity of the area can be organized by data modality and inference objective.

Modality	Typical target	Representative studies
Names	Individual label or group composition	(Hu et al., 2021, Antonoyiannakis et al., 2023, Buskirk et al., 2022)
Visual and biometric imagery	Binary or ternary classification; visible representation	(Kuehlkamp et al., 2017, Kuehlkamp et al., 2018, Wu et al., 2020, Mazieres et al., 2020)
Text and dialogue	Gendered language, character portrayal, system ontology	(Hoyle et al., 2019, Keith et al., 2024, Hafner et al., 20 May 2025)
Behavior and networks	Gender from mobility, personality, interaction structure	(Psylla et al., 2017)
Health and administrative records	Gender-related status from proxies	(Lokala et al., 2022, Gronsbell et al., 19 Aug 2025, Ogundare et al., 2024)

In name-based systems, the simplest pipeline maps names to probabilities or labels using reference distributions, APIs, or historical registries. More elaborate approaches use character n-grams with tf-idf and log-ratio scaling, LSTMs, byte-level CNNs, or transformer variants (Hu et al., 2021). Ensemble methods integrate multiple public name-gender sources through a Cultural Consensus Theory-inspired EM procedure, estimating both source competence and a latent consensus label for each name (Buskirk et al., 2022). The global gender estimation method replaces independent per-name assignment with a self-consistent population-level equation over the full name distribution (Antonoyiannakis et al., 2023).

Text-based systems span several distinct designs. One class uses generative latent-variable models over nouns, modifiers, and latent sentiment, as in

$p(\nu, n, s)=p(\nu \mid s,n)\,p(s\mid n)\,p(n),$

to discover gendered collocational patterns in large corpora (Hoyle et al., 2019). Another fine-tunes pretrained contextual LLMs such as BETO for binary character-speech classification in literary corpora, then uses integrated gradients to expose token-level cues (Keith et al., 2024). A third uses prompt-based conditional probability probing of pretrained LLMs, computing log-probability ratios over controlled templates to audit implicit gender construction (Hafner et al., 20 May 2025).

Visual phenotyping includes face-region pipelines, CNN-based facial classifiers, iris and periocular classification, and large-scale movie analysis based on face detection and automated gender inference. The movie pipeline samples frames uniformly every 2 seconds, applies Wolfram Mathematica Engine 12’s FacialFeatures, validates outputs by manual review, and corrects the raw female face ratio by the confusion matrix because female predictions are overestimated by the classifier (Mazieres et al., 2020). Biometric studies likewise emphasize that ROI choice matters: normalized iris, periocular region, segmentation masks, and whole-eye imagery can carry materially different signals (Kuehlkamp et al., 2018).

Behavioral phenotyping uses multi-channel digital traces. In the Copenhagen Network Study, weekly aggregated questionnaire, mobility, Facebook, Bluetooth proximity, and phone metadata are combined into feature vectors and classified with Logistic Regression, AdaBoost, SVC, Random Forest, and Gradient Boosting under 10-fold stratified cross-validation (Psylla et al., 2017). Health-related text phenotyping employs multitask setups; GeM uses two RoBERTa-based encoders, ontology- and lexicon-based masking, and joint learning of symptom identification and gender identification in Reddit posts about cardiovascular disease (Lokala et al., 2022).

3. Name-based phenotyping, population estimation, and temporal instability

Name-based gender inference is one of the most mature forms of computational phenotyping because names are nearly ubiquitous in administrative and bibliometric records. On a Verizon Media / Yahoo dataset of 100M+ users and 21M unique first names, character-based models substantially outperform a content baseline: NBLR reaches AUC $= 0.940$, ACC $= 0.876$ on the full Yahoo test set, while LSTM and char-BERT are similar, and adding last names via DLSTM improves to AUC $= 0.957$, ACC $= 0.901$ on Yahoo Full Name Data (Hu et al., 2021). The same work reports that prefixing and suffixing each token with “_” yields an additional 1–3% AUC improvement, and that character models generalize better than embedding methods to rare names, where NBLR/LSTM/char-BERT remain around 0.872–0.874 AUC while the embedding baseline falls to about 0.724 AUC (Hu et al., 2021).

Yet several papers show that name-based phenotyping is fragile when context is ignored. The global gender estimation paper argues that conventional individual-based gender estimation methods (iGEMs) are logically inconsistent for skewed populations because they assume $p_R(g\mid s)\approx p_T(g\mid s)$. Its alternative, gGEM, estimates group composition self-consistently using

$$\sum_{s} \frac{\delta_R(s)}{1+\delta_R(s)\gamma}N(s)=0,$$

where $\delta_R(s)=p_R(f\mid s)-p_R(m\mid s)$, and shows that iGEMs systematically underestimate gender bias by overestimating minority-gender participation (Antonoyiannakis et al., 2023). This redefines phenotyping from independent label assignment to global latent-variable estimation.

Open-source ensemble methods address a different problem: fragmented coverage and lack of standardization. The Cultural Consensus Theory approach combines 36 public data sources spanning 150+ countries, more than a century, and over 500,000 unique names. It reports about 96% overall correspondence with validation labels, with performance depending strongly on name type: about 99% for high-coverage informative names, about 93% for low-coverage informative names, about 70% for weakly gendered names, and variable performance for conditionally gendered names (Buskirk et al., 2022). The same paper argues that many systems are already near the practical ceiling imposed by ambiguity, Romanization, and missing coverage.

Historical work shows that temporal instability is an equally serious problem. Misa’s “Leslie problem” demonstrates that many names change gender association over decades. Using the SSA dataset, he computes $$\{a\ man, a\ woman, transgender, nonbinary, genderqueer, genderfluid, two\text{-}spirit\}$$0 by year and identifies more than 300 names with measurable shifts across 1925–2000, with a median shift of +55.36 when $$\{a\ man, a\ woman, transgender, nonbinary, genderqueer, genderfluid, two\text{-}spirit\}$$1 is multiplied by 100 (Misa, 2022). Names such as Shelby, Leslie, Aubrey, and Sydney are central examples. A companion analysis comparing Gender API, NamSor, and Genderizer.io against personally identified DBLP authors from 1950–1980 reports a median of the diverging ratios of 2.13, a mean of 14.6, and individual inflated ratios ranging from 25-fold to 70-fold, indicating substantial type-two bias in which historically male authors are misidentified as female (Misa, 2022).

Cross-national evaluation reveals further heterogeneity. A comparative study on 1,416 scientists finds overall accuracies of 0.75 for SSA, 0.68 for IPUMS, 0.74 for Sexmachine, 0.82 for Genderize, 0.83 for Face++, 0.92 for Mixed1, and 0.91 for Mixed2, but the strongest bias is by country rather than by male/female class (Karimi et al., 2016). For China, for example, SSA reaches 0.20 and IPUMS 0.11, while Face++ reaches 0.65; for South Korea, SSA is 0.04, IPUMS 0.00, Sexmachine 0.58, Genderize 0.11, and Face++ 0.74 (Karimi et al., 2016). The combined literature therefore establishes three constraints on name-based phenotyping: cultural coverage, temporal alignment, and the distinction between individual labeling and population composition.

4. Vision, biometrics, and visible representation

Visual phenotyping studies repeatedly show that the apparent “gender signal” depends on what part of the image is actually being measured. In iris biometrics, a stricter experimental protocol sharply reduces reported performance. Under person-disjoint train/test splits averaged over 10 randomized trials, the original GFI iris study reports $$\{a\ man, a\ woman, transgender, nonbinary, genderqueer, genderfluid, two\text{-}spirit\}$$2, compared with $$\{a\ man, a\ woman, transgender, nonbinary, genderqueer, genderfluid, two\text{-}spirit\}$$3 on a non-person-disjoint dataset, and notes that a single non-person-disjoint split can yield 100% accuracy (Kuehlkamp et al., 2017). The same study shows that just over 60% of female images exhibit cosmetics while 0% of male images do, that average image intensity alone separates Males vs FWC at better than 60% accuracy but Males vs FNC at only about 50%, and that classification using only the binary occlusion mask is similar to using iris imagery, indicating substantial confounding by mascara, eyelash occlusion, and imperfect segmentation (Kuehlkamp et al., 2017).

A follow-up study on the larger GFI-C dataset reaches the same substantive conclusion. On normalized iris images, handcrafted features + linear SVM yield $$\{a\ man, a\ woman, transgender, nonbinary, genderqueer, genderfluid, two\text{-}spirit\}$$4, VGG features + linear SVM yield $$\{a\ man, a\ woman, transgender, nonbinary, genderqueer, genderfluid, two\text{-}spirit\}$$5, and fine-tuned VGGNet-16 reaches $$\{a\ man, a\ woman, transgender, nonbinary, genderqueer, genderfluid, two\text{-}spirit\}$$6. On periocular images, by contrast, performance jumps to $$\{a\ man, a\ woman, transgender, nonbinary, genderqueer, genderfluid, two\text{-}spirit\}$$7 and $$\{a\ man, a\ woman, transgender, nonbinary, genderqueer, genderfluid, two\text{-}spirit\}$$8 (Kuehlkamp et al., 2018). Probabilistic occlusion masking then drives normalized-image performance from about $$\{a\ man, a\ woman, transgender, nonbinary, genderqueer, genderfluid, two\text{-}spirit\}$$9 at threshold 0 down to $p(\nu, n, s)=p(\nu \mid s,n)\,p(s\mid n)\,p(n),$0 when only pixels with $p(\nu, n, s)=p(\nu \mid s,n)\,p(s\mid n)\,p(n),$1 non-occlusion are retained, showing that the strongest cues are near the iris rather than in the iris stroma itself (Kuehlkamp et al., 2018).

Face-based classification shows a different but related problem: benchmark scope. A binary CNN trained on Adience achieves 94.37% accuracy on the Adience test set, but when applied to a new inclusive/non-binary test set of 3,000 images, accuracy drops to 51.67%, with 0% accuracy for the non-binary class (Wu et al., 2020). After constructing an inclusive benchmark database and a non-binary benchmark database, transfer learning and ensemble methods markedly reduce disparity; the best logistic regression ensemble reaches 90.39% overall, with subgroup accuracies of 90.02% for male, 88.65% for female, and 91.97% for non-binary, and a selection rate of 96.40% (Wu et al., 2020). Region-aware classical pipelines arrive at a complementary result: in frontal facial images, the eye regions carry the most gender-specific information, with Adience region-wise accuracies of 84.06% for the left eye and 82.93% for the right eye, while the forehead is least informative at 73.95% (Bhattacharyya et al., 2017).

A distinct visual line of work treats gender not as an attribute of a depicted person but as a property of cultural representation. In a corpus of 3,776 movies sampled from 1985–2019, frames are extracted every 2 seconds, yielding more than 12.4 million images, and passed through automated face and gender detection (Mazieres et al., 2020). Manual validation reports face detection accuracy of 92% and gender inference accuracy of 73.9% overall, with a marked asymmetry: predicted-female labels are correct 65% of the time, predicted-male labels 84.5% (Mazieres et al., 2020). After confusion-matrix correction, the study uses the female face ratio to show that mean visible female presence rises from about 27% in 1985–1998 to 44.9% in 2014–2019, while genre, economic, and spatial patterns remain structured (Mazieres et al., 2020). The common methodological lesson across these studies is that visual gender phenotypes are often phenotypes of presentation, context, or framing rather than of an isolated anatomical substrate.

5. Language, behavior, and institutional phenotypes

Text-centered work broadens computational phenotyping from visible cues to linguistic, semantic, and conceptual structure. The audit of 16 open-source pretrained LLMs is exemplary in this regard. Using prompt-based conditional probability probing over controlled templates, it defines the log probability ratio (LPR),

$p(\nu, n, s)=p(\nu \mid s,n)\,p(s\mid n)\,p(n),$2

and evaluates Sex–Gender and Gender–Illness associations (Hafner et al., 20 May 2025). The reported findings are that all models encode gender as a binary category tied to biological sex; larger models intensify this alignment, with Folk–Subversive LPR rising from 0.42 for T5-small to 3.22 for Mixtral-8x7B-v0.1, and a Spearman rank correlation of $p(\nu, n, s)=p(\nu \mid s,n)\,p(s\mid n)\,p(n),$3 between size and Folk–Subversive LPR (Hafner et al., 20 May 2025). The same audit shows that terms such as nonbinary, transgender, genderqueer, genderfluid, and two-spirit are often less probable than a baseline of 47 non-human random nouns, and that gender-diverse contexts shift completion distributions toward mental illness terms such as body dysmorphia, post-traumatic stress disorder, panic disorder, and anorexia nervosa (Hafner et al., 20 May 2025).

Other language studies phenotype gendered discourse directly. The latent-variable model for adjective and verb collocations over an 11 billion word corpus finds that there are significant differences in how male and female nouns are described, and that positive adjectives used to describe women are more often body-related than positive adjectives used to describe men, with $p(\nu, n, s)=p(\nu \mid s,n)\,p(s\mid n)\,p(n),$4 under permutation testing (Hoyle et al., 2019). In literary analysis, a BETO-based classifier trained on 109 Calderón comedias reaches macro F1 = 0.83 at the character level, while scene- and utterance-level predictions improve substantially when aggregated by the geometric mean of probabilities rather than majority vote, achieving F1 0.81 and 0.76 respectively (Keith et al., 2024). Because scene-by-scene predictions often flip during disguise episodes, the model provides a relatively accurate dynamic signal for cross-dressed characters such as Rosaura and Semíramis (Keith et al., 2024).

Behavioral phenotyping moves the locus of measurement from language to social organization. In a cohort of 166 females and 601 males from the Copenhagen Network Study, the strongest classifiers achieve approximately ROC-AUC $p(\nu, n, s)=p(\nu \mid s,n)\,p(s\mid n)\,p(n),$5 and $p(\nu, n, s)=p(\nu \mid s,n)\,p(s\mid n)\,p(n),$6, with network features dominating importance: five of the top six indicators come from homophily and communication structure (Psylla et al., 2017). Women in this cohort show stronger same-gender homophily, more same-gender triangles, higher interaction entropy in several channels, more text messages, and longer phone calls (Psylla et al., 2017). The same paper interprets these predictions explicitly as behavioral correlates rather than an intrinsic essence of gender.

Health-related social media work further entangles gender with another latent phenotype. The GeM framework uses a knowledge-assisted, task-adaptive, RoBERTa-based bi-encoder multitask model to jointly classify four mental-health symptoms and two gender labels from Reddit posts/comments about cardiovascular disease (Lokala et al., 2022). Built from a raw crawl of 150k Reddit items, then filtered to a CVDS corpus of about 16k posts/comments from about 5,400 users, it improves over RoBERTa by 2.14% recall on symptom identification and 2.55% recall on gender identification (Lokala et al., 2022). This line of work treats gender not only as a target label but also as a conditioning variable that changes how symptom language is expressed.

Institutional measurement can also be framed as phenotyping. The Sarafina score models the gender asset gap in land acquisition and land tenure security as a dynamic state inferred from nominal ownership disparity and expected policy impact (Ogundare et al., 2024). It defines

$p(\nu, n, s)=p(\nu \mid s,n)\,p(s\mid n)\,p(n),$7

with policy impact estimated from proxy indicators including Economic GDP, Higher education gender ratio, Birth rate, Domestic violence incidence–investigation ratio, and Judicial effectiveness (Ogundare et al., 2024). A plausible implication is that computational phenotyping of gender includes not only person-level classification but also policy-sensitive measurement of gendered institutional conditions.

6. Validation, bias, and ethical controversy

Across modalities, the dominant methodological theme is that computational gender phenotypes are highly sensitive to confounding, miscalibration, and mislabeled ground truth. Iris studies show inflation from non-person-disjoint splits and from cosmetics leaking into masks or whole-eye images (Kuehlkamp et al., 2017). Movie analysis explicitly measures model error on the target domain and corrects the female face ratio because the classifier overestimates female faces and does so more strongly in earlier periods (Mazieres et al., 2020). EHR work is even more explicit: performance metrics such as TPR, FPR, PPV, NPV, F1, AUC, and AUPRC can be misleading because the gold standard is usually chart review, which is itself an imperfect proxy for gender and often identifies only well-documented, medicalized cases (Gronsbell et al., 19 Aug 2025).

Bias also enters through operational closure. Binary models can erase non-binary, transgender, and gender-diverse identities. The language-model audit reports erasure and pathologization of nonbinary, transgender, genderqueer, genderfluid, and two-spirit terms (Hafner et al., 20 May 2025). The facial-image inclusion study shows that a binary classifier trained on Adience assigns 0% accuracy to the non-binary class when transferred to an inclusive benchmark (Wu et al., 2020). EHR review articles argue that many phenotypes collapse gender identity, gender expression, and medical transition into a single inferred label and thereby reproduce outdated or pathologizing notions of gender (Gronsbell et al., 19 Aug 2025).

Historical and cultural mismatch are equally consequential. The “Leslie problem” demonstrates that present-day name-gender associations can systematically distort historical analyses (Misa, 2022). Country-level evaluation shows that name-only methods work far better for Western industrialized countries than for China, South Korea, Brazil, India, and Turkey, and that hybrid image-plus-name methods partly reduce but do not eliminate this bias (Karimi et al., 2016). Open-source consensus methods find that Romanized Chinese names are disproportionately concentrated in no-data or weakly gendered categories, indicating a structural ceiling on inference quality for some populations (Buskirk et al., 2022).

The ethical stakes differ by domain but are non-trivial in all of them. The language-model audit links narrow sex-gender associations to downstream harms such as misgendering in chatbots and misdiagnosis in healthcare, including diagnostic overshadowing (Hafner et al., 20 May 2025). EHR critiques warn that computational phenotyping can shift from inclusion to surveillance, especially in environments marked by criminalization of gender-affirming care, and emphasize risks of re-identification, misclassification, and harm to trans and gender-expansive communities (Gronsbell et al., 19 Aug 2025). Name-based researchers caution that misgendering is a real risk and that inferred labels should not be treated as substitutes for self-reported identity or pronouns (Buskirk et al., 2022).

Several convergent best-practice directions emerge. Studies recommend person-disjoint evaluation, manual validation on the target domain, subgroup-aware reporting, and explicit correction for classifier asymmetry where possible (Kuehlkamp et al., 2017). Name-based work recommends auditing country, decade, and name-type composition rather than reporting only overall accuracy (Buskirk et al., 2022). Language-model auditing recommends theory-informed definitions of gender, broader representation of gender diversity in training data, and avoidance of debiasing strategies that merely fix binary associations while erasing other identities (Hafner et al., 20 May 2025). EHR critiques add a more fundamental recommendation: move away from default gender assignment, specify exactly what construct is being measured, adopt richer concepts such as gender modality, and center community accountability and self-determination in study design (Gronsbell et al., 19 Aug 2025).

Taken together, these studies show that computational phenotyping of gender is best understood as a measurement problem under proxy uncertainty. Its technical forms are diverse—classification, ranking, latent-variable estimation, rule-based heuristics, ensemble consensus, and critical system audit—but its recurrent scientific questions are stable: what is being measured, from which signals, under which assumptions, with which sources of bias, and to what downstream end.