CIFID: Clinically-Informed Feature Dictionary

• CIFID is a structured interpretability framework that converts latent model features into clinically meaningful labels through expert-curated and sparse dictionary methods.
• It supports diverse modalities—including radiomics, pathomics, and medical coding—by aligning features with standardized clinical ontologies like PI-RADS and ICD.
• The framework enables transparent model interpretation by providing formal feature representations that facilitate verification, reproducibility, and actionable clinical insights.

Searching arXiv for the cited CIFID-related papers and closely related work. Clinically-Informed Feature Interpretation Dictionary (CIFID) denotes a family of structured interpretability frameworks that translate quantitative or latent model features into clinically meaningful semantics. In the current literature, CIFID appears in at least two recurrent forms: a sparse, monosemantic dictionary learned from neural representations, and an expert-curated dictionary aligning IBSI-compliant radiomic or pathomic features with clinical ontologies such as PI-RADS, BI-RADS, TI-RADS, Lung-RADS, and WHO grading. Across these realizations, the central operation is the same: a model feature is assigned a formal representation, a semantic label, and an auditable relation to prediction, so that explanations can be inspected in domain language rather than inferred from opaque embeddings alone (Tang et al., 24 Oct 2025, Salmanpour et al., 2024, Salmanpour et al., 20 May 2025, Wu et al., 2024).

1. Conceptual scope and major variants

Published CIFID realizations span chest radiography, prostate MRI, thyroid ultrasound, lung CT, breast DCE-MRI, liver histopathology, automated medical coding, and clinical LLM attribution. The dictionary unit varies by domain: it may be a discovered visual pattern, an IBSI feature row, a sparse dictionary atom, or an SAE feature. What remains consistent is that each entry is designed to function as a clinically grounded intermediary between model internals and human interpretation (Tang et al., 24 Oct 2025, Salmanpour et al., 24 Mar 2026, Jouzdani et al., 31 Dec 2025, Gorji et al., 21 Jul 2025, Wu et al., 2024, Mamalakis et al., 25 Jan 2026).

Domain	Dictionary unit	Clinical grounding
Chest X-ray	Monosemantic visual pattern	Cardiac, pulmonary, pleural, structural, device, artifact
Prostate MRI PM1.0	Radiomic feature entry	PI-RADS or risk factors
Liver cancer LCP1.0	Tabular imaging-feature row	WHO Grade Mapping
Thyroid US TU1.0	IBSI radiomic feature	TI-RADS descriptors
Lung CT LC 1.0	IBSI radiomic feature	Lung-RADS descriptors
Breast MRI BM1.0	Radiomic feature mapping	BI-RADS descriptors
Medical coding	Sparse dictionary feature / atom	ICD-code associations
Clinical neuroscience LLMs	SAE feature entry	Biomarker- or token-level salience

A common misconception is that CIFID denotes only one technical recipe. The literature instead uses the term for a broader interpretability pattern. In imaging radiomics and pathomics, CIFID is frequently ontology-centered and expert validated. In sparse-representation work on PLMs or multimodal encoders, CIFID is instead representation-centered, with semantics grounded post hoc through activation galleries, ablation, LLM naming, or downstream code effects. This suggests that CIFID is better understood as a design principle for clinically anchored feature semantics than as a single algorithm.

2. Formal representations and learning objectives

In the chest X-ray realization built on CXR-LanIC, CIFID is explicitly a dictionary over discovered monosemantic patterns. Let $P=\{p_1,\dots,p_N\}$ denote the set of curated patterns with $N\approx 5{,}000$. Every CXR image $x$ is mapped to a sparse activation vector $z(x)\in\mathbb{R}^N$. Each entry has the form

$$\mathrm{CIFID}_p \triangleq (\mathrm{pattern\_id}=p,\ \mathrm{desc}_p,\ \mathrm{cat}_p,\ \tau_p,\ \mathrm{Ex}_p),$$

where $$\mathrm{cat}_p\in\{\mathrm{cardiac},\mathrm{pulmonary},\mathrm{pleural},\mathrm{structural},\mathrm{device},\mathrm{artifact}\}$$. The underlying transcoder ensemble uses $100$ transcoders with $d=1024$, $m=15{,}000$, and Top-$K$ sparsification with $N\approx 5{,}000$0; for transcoder $N\approx 5{,}000$1,

$N\approx 5{,}000$2

with loss

$N\approx 5{,}000$3

After merging $N\approx 5{,}000$4M neurons, the top $N\approx 5{,}000$5 clinically consistent neurons are retained, pattern activation is defined as

$N\approx 5{,}000$6

and the thresholded feature value is

$N\approx 5{,}000$7

This formulation makes CIFID simultaneously a sparse code, a thresholded detector, and a semantic index over activation exemplars (Tang et al., 24 Oct 2025).

In medical coding, CIFID is instantiated as a sparse dictionary over PLM token embeddings. Both the dictionary-learning framework for coding and DILA use an autoencoding decomposition of dense embeddings into sparse features. A representative formulation is

$N\approx 5{,}000$8

optimized with reconstruction plus sparsity penalties such as

$N\approx 5{,}000$9

DILA then replaces conventional nonlinear label attention with a single interpretable matrix $x$0 linking dictionary features directly to ICD codes. In that model, $x$1 active features are reported, and only $x$2 of a token’s $x$3k features are nonzero (Wu et al., 2024, Wu et al., 2024).

In the monosemantic attribution framework for clinical neuroscience LLMs, CIFID is anchored in an SAE bottleneck over a chosen transformer layer. The layer activation $x$4 is decomposed as

$x$5

and the explanation module is optimized with a stability-aware objective

$x$6

This makes the dictionary not only interpretable but explicitly optimized for inter-method stability (Mamalakis et al., 25 Jan 2026).

3. Construction workflows

The learned-dictionary workflow is especially explicit in CXR-LanIC. For each transcoder, a random $x$7 subset of images and associated report embeddings is drawn; optional random linear projection on $x$8 is used to encourage diversity; different seeds initialize each $x$9; and only neurons that consistently reconstruct $z(x)\in\mathbb{R}^N$0 across $z(x)\in\mathbb{R}^N$1 of held-out splits are retained as pattern candidates. Candidate neurons are then assigned semantics through a multistage procedure: compute activations over the training set, extract Top-$z(x)\in\mathbb{R}^N$2 high-activation images, prompt Claude-4.5 with 10 CXR crops and report snippets for a shared finding, cluster descriptions into six categories, and retain only patterns with $z(x)\in\mathbb{R}^N$3 and activation frequency $z(x)\in\mathbb{R}^N$4. Verification uses 128×128 pixel crops centered at highest-activation receptive fields and the purity statistic

$z(x)\in\mathbb{R}^N$5

with only patterns satisfying $z(x)\in\mathbb{R}^N$6 preserved (Tang et al., 24 Oct 2025).

In radiomics and pathomics dictionaries, the workflow is usually ontology first and model second. LCP1.0 extracts $z(x)\in\mathbb{R}^N$7 features from TCGA-LIHC H&E whole-slide ROIs using QuPath v0.2.3 and PyRadiomics v3.0, both standardized per IBSI guidelines, and then organizes each dictionary row by name, source, units, extraction parameters, definition, pathobiological interpretation, and WHO Grade Mapping. BM1.0 begins from $z(x)\in\mathbb{R}^N$8 PyRadiomics v3.0 features, uses literature-driven descriptor identification and optional Spearman pre-filtering, then applies a 2-round Delphi process to accept only RF-to-descriptor mappings with sufficient agreement. TU1.0 extracts $z(x)\in\mathbb{R}^N$9 radiomic features from three multicenter ultrasound cohorts, retains $$\mathrm{CIFID}_p \triangleq (\mathrm{pattern\_id}=p,\ \mathrm{desc}_p,\ \mathrm{cat}_p,\ \tau_p,\ \mathrm{Ex}_p),$$0 common features for modeling, and uses expert consensus plus SHAP to validate mappings. LC 1.0 similarly starts from $$\mathrm{CIFID}_p \triangleq (\mathrm{pattern\_id}=p,\ \mathrm{desc}_p,\ \mathrm{cat}_p,\ \tau_p,\ \mathrm{Ex}_p),$$1 IBSI-compliant radiomic features and uses independent expert assignment followed by consensus meeting and majority vote for feature–descriptor pairing (Salmanpour et al., 20 May 2025, Gorji et al., 21 Jul 2025, Salmanpour et al., 24 Mar 2026, Jouzdani et al., 31 Dec 2025).

In coding and LLM settings, the workflow shifts from clinical lexicon curation to mechanistic grounding. In the coding CIFID derived from dictionary learning, an atom is interpreted by ablating an activated feature $$\mathrm{CIFID}_p \triangleq (\mathrm{pattern\_id}=p,\ \mathrm{desc}_p,\ \mathrm{cat}_p,\ \tau_p,\ \mathrm{Ex}_p),$$2 from token embedding $$\mathrm{CIFID}_p \triangleq (\mathrm{pattern\_id}=p,\ \mathrm{desc}_p,\ \mathrm{cat}_p,\ \tau_p,\ \mathrm{Ex}_p),$$3, recomputing the model, and ranking ICD codes by average probability drop $$\mathrm{CIFID}_p \triangleq (\mathrm{pattern\_id}=p,\ \mathrm{desc}_p,\ \mathrm{cat}_p,\ \tau_p,\ \mathrm{Ex}_p),$$4. DILA augments this with an LLM-based naming pipeline: top activating contexts are subjected to an odd-one-out coherence test, then summarized into an 8-word description. This yields a scalable, weakly supervised route to thousands of feature names without manual annotation (Wu et al., 2024, Wu et al., 2024).

4. Clinical semantics and ontology alignment

The ontology-centered CIFID literature is organized around established reporting systems. In prostate MRI PM1.0, five key features are defined with mathematical formulas, acquisition sequence, interpretation guidance, and example thresholds. FO_90P captures the brightest 10% of lesion intensity on T2WI; lower values correspond to hypo-intense regions characteristic of PI-RADS 4–5, with an example threshold for high risk of $$\mathrm{CIFID}_p \triangleq (\mathrm{pattern\_id}=p,\ \mathrm{desc}_p,\ \mathrm{cat}_p,\ \tau_p,\ \mathrm{Ex}_p),$$5. FO_V quantifies intensity dispersion on T2WI, with higher values indicating heterogeneity and an example high-risk threshold $$\mathrm{CIFID}_p \triangleq (\mathrm{pattern\_id}=p,\ \mathrm{desc}_p,\ \mathrm{cat}_p,\ \tau_p,\ \mathrm{Ex}_p),$$6. ADC-derived shape features include $$\mathrm{CIFID}_p \triangleq (\mathrm{pattern\_id}=p,\ \mathrm{desc}_p,\ \mathrm{cat}_p,\ \tau_p,\ \mathrm{Ex}_p),$$7, where higher values suggest bulkier or spherical lesions, and $$\mathrm{CIFID}_p \triangleq (\mathrm{pattern\_id}=p,\ \mathrm{desc}_p,\ \mathrm{cat}_p,\ \tau_p,\ \mathrm{Ex}_p),$$8, where higher values indicate irregularity. GLRLM_REn quantifies randomness in run-length distributions, with higher entropy linked to heterogeneous malignant tissue and an example threshold $$\mathrm{CIFID}_p \triangleq (\mathrm{pattern\_id}=p,\ \mathrm{desc}_p,\ \mathrm{cat}_p,\ \tau_p,\ \mathrm{Ex}_p),$$9 (Salmanpour et al., 2024).

TI-RADS-based TU1.0 maps the five thyroid semantic categories—composition, echogenicity, shape, margin, and echogenic foci—to IBSI features with explicit interpretation rules. Examples include “If Elongation $$\mathrm{cat}_p\in\{\mathrm{cardiac},\mathrm{pulmonary},\mathrm{pleural},\mathrm{structural},\mathrm{device},\mathrm{artifact}\}$$0, then map to ‘taller-than-wide’,” $$\mathrm{cat}_p\in\{\mathrm{cardiac},\mathrm{pulmonary},\mathrm{pleural},\mathrm{structural},\mathrm{device},\mathrm{artifact}\}$$1, $$\mathrm{cat}_p\in\{\mathrm{cardiac},\mathrm{pulmonary},\mathrm{pleural},\mathrm{structural},\mathrm{device},\mathrm{artifact}\}$$2, and $$\mathrm{cat}_p\in\{\mathrm{cardiac},\mathrm{pulmonary},\mathrm{pleural},\mathrm{structural},\mathrm{device},\mathrm{artifact}\}$$3 for hypoechoic or very hypoechoic nodules, and $$\mathrm{cat}_p\in\{\mathrm{cardiac},\mathrm{pulmonary},\mathrm{pleural},\mathrm{structural},\mathrm{device},\mathrm{artifact}\}$$4 with elevated LoG response for punctate echogenic foci (Salmanpour et al., 24 Mar 2026).

Lung-RADS-based LC 1.0 extends the same logic to ten pulmonary nodule descriptors, including shape, margin, attenuation, growth pattern, calcification, internal features, location, spiculation, and associated findings. The mapped examples reported in the dictionary include $$\mathrm{cat}_p\in\{\mathrm{cardiac},\mathrm{pulmonary},\mathrm{pleural},\mathrm{structural},\mathrm{device},\mathrm{artifact}\}$$5 margin irregularity or spiculation, $$\mathrm{cat}_p\in\{\mathrm{cardiac},\mathrm{pulmonary},\mathrm{pleural},\mathrm{structural},\mathrm{device},\mathrm{artifact}\}$$6 irregular shape, $$\mathrm{cat}_p\in\{\mathrm{cardiac},\mathrm{pulmonary},\mathrm{pleural},\mathrm{structural},\mathrm{device},\mathrm{artifact}\}$$7 solid attenuation, $$\mathrm{cat}_p\in\{\mathrm{cardiac},\mathrm{pulmonary},\mathrm{pleural},\mathrm{structural},\mathrm{device},\mathrm{artifact}\}$$8 fine spiculation, and $$\mathrm{cat}_p\in\{\mathrm{cardiac},\mathrm{pulmonary},\mathrm{pleural},\mathrm{structural},\mathrm{device},\mathrm{artifact}\}$$9 heterogeneity or associated findings (Jouzdani et al., 31 Dec 2025).

BI-RADS-centered BM1.0 uses a dual dictionary. CIFID maps 56 RFs to Shape, Margin, and Internal Enhancement descriptors, while DDFID provides SHAP-derived tags for 52 additional RFs. Representative shape mappings include Sphericity, Flatness, Perimeter-to-Surface Ratio, and Surface-Area-to-Volume Ratio; homogeneous internal enhancement is associated with features such as Uniformity and lower Busyness (Gorji et al., 21 Jul 2025).

In CXR-LanIC, semantics are not inherited from a formal radiology ontology but discovered as monosemantic visual patterns spanning cardiac, pulmonary, pleural, structural, device, and artifact categories. An example entry is pattern P0427, category pleural, description “Right pleural effusion,” with threshold $100$0 and top exemplar crops. This indicates that CIFID can be ontology-aligned either by direct mapping to standardized lexicons or by clinically grounded naming of recurring latent patterns (Tang et al., 24 Oct 2025).

5. Explanation generation, evaluation, and validation

In the chest X-ray setting, CIFID generates natural-language explanations directly from the sparse pattern vector $100$1. Given a test image, the active set $100$2 is sorted by descending activation, the top $100$3 patterns are selected, and a template yields findings of the form: pattern description, pattern identifier, and score, followed by a diagnostic suggestion. The specification instructs that only patterns with $100$4 be included, that patterns be grouped by anatomical category for readability, and that the final suggestion be based on combination rules such as enlarged heart plus bilateral interstitial changes plus effusion implying CHF. In the broader CXR-LanIC system, predictions decompose into 20–50 interpretable patterns with verifiable activation galleries, and the framework reports competitive diagnostic accuracy on five key findings (Tang et al., 24 Oct 2025).

In automated medical coding, CIFID explanations are mechanistic rather than merely local. The coding dictionary-learning formulation explains a prediction through the chain “token embedding $100$5 sparse dictionary feature $100$6 $100$7 downstream effect on code $100$8,” or equivalently $100$9. Quantitatively, L1-AE CIFID (LAAT + DL) reports $d=1024$0, $d=1024$1, and $d=1024$2 on comprehensiveness; among 12,891 LAAT-highlighted stop words, it recovers the original ICD code in the top-10 per-feature list 91% of the time; and model steering by clamping a single atom to $d=1024$3 on blank PAD input can flip up to 3,681 ICD codes with 0.89 accuracy (Wu et al., 2024). DILA complements this with global transparency: sparse embeddings are reported as more human understandable than dense counterparts by at least 50 percent; 3,524 out of 6,088 dictionary features are identifiable by LLM; and the model maintains competitive end-to-end performance on MIMIC-III clean, including Micro-F1 54.9% and Micro AUC-ROC 97.6% (Wu et al., 2024).

In ontology-centered radiomics and pathomics, validation is usually a combination of predictive performance, SHAP consistency, and expert review. PM1.0 reports highest average accuracy of 0.78 with multisequence logistic regression. LCP1.0 reports mean cross-validation accuracy $d=1024$4 and external test accuracy $d=1024$5, with paired $d=1024$6-test $d=1024$7. TU1.0 reports a hold-out test ROC-AUC of $d=1024$8, with SHAP showing texture heterogeneity as the dominant malignancy signal. LC 1.0 reports mean validation accuracy 0.79 and inter-rater agreement $d=1024$9 for feature–descriptor assignments. BM1.0 reports average cross-validation accuracy 0.83 and average Delphi agreement $m=15{,}000$0 (Salmanpour et al., 2024, Salmanpour et al., 20 May 2025, Salmanpour et al., 24 Mar 2026, Jouzdani et al., 31 Dec 2025, Gorji et al., 21 Jul 2025).

These validation protocols show that CIFID is not restricted to one notion of interpretability. In some works, the relevant criterion is activation purity; in others, it is expert agreement, SHAP alignment with a clinical lexicon, comprehensiveness under ablation, or stability under attributional perturbation. A plausible implication is that CIFID should be assessed relative to the explanatory contract of the host model rather than by a single universal metric.

6. Limitations, misconceptions, and prospective directions

The literature repeatedly emphasizes that CIFID does not remove all ambiguity from model interpretation. In coding dictionaries, reconstruction is imperfect, some features remain dead, granularity is limited because a few thousand atoms must cover thousands of ICD codes, and scaling to larger models may require millions of atoms. DILA further notes that not all dictionary features are identifiable, that LLM summaries can be too general or slightly wrong, that rare clinical abbreviations may never surface enough contexts to form a coherent feature, and that direct edits to $m=15{,}000$1 can reduce false positives for one code while introducing false negatives elsewhere (Wu et al., 2024, Wu et al., 2024).

In radiomics and pathomics, the primary limitations concern reproducibility and semantic fidelity. TU1.0 requires standardized segmentation, training-fold-only min–max normalization, fixed random seeds, and documentation of ultrasound acquisition parameters. LCP1.0 recommends full documentation of formulas, units, imaging parameters, version control of the dictionary, audit trails for expert review, adoption of standard vocabularies such as SNOMED CT and RadLex, and early engagement with regulatory bodies for software-as-a-medical-device submissions (Salmanpour et al., 24 Mar 2026, Salmanpour et al., 20 May 2025).

In monosemantic feature discovery for clinical LLMs, stability itself becomes a maintenance target. The neuroscience framework recommends periodic recomputation of attributions on new data, monitoring RIS and ROS over time, updating feature descriptions when clinical biomarkers evolve, and version-controlling entries by date, model version, and SAE variant (Mamalakis et al., 25 Jan 2026).

Another misconception is that CIFID is equivalent to post-hoc highlighting. The cited work shows a broader design space: CIFID may be a lookup table over handcrafted IBSI features, a sparse code over PLM tokens, a monosemantic SAE basis, or an ensemble-derived visual pattern dictionary. Future directions stated in the literature include planned large multimodal model annotation for chest X-ray patterns, richer disentangled or causal representation learning for coding, automated circuit discovery alongside dictionary atoms, extension to 3D US and shear-wave elastography, and re-anchoring the same framework to other ontologies such as PI-RADS or BI-RADS in new organ systems (Tang et al., 24 Oct 2025, Wu et al., 2024, Salmanpour et al., 24 Mar 2026, Jouzdani et al., 31 Dec 2025).

Taken together, CIFID has emerged as a unifying interpretability template for clinically grounded AI: a dictionary layer or dictionary table that makes features inspectable, semantically named, and operationally connected to prediction. Its importance lies less in any single implementation than in the recurring insistence that clinically deployable models should expose their decisive features in the language of the domain, with sufficient formal structure to support verification, comparison, and revision.