Bone is constantly rebuilt by two cell types working against each other. Osteoclasts dissolve old bone, and osteoblasts lay down new bone. Many treatments, drugs and metabolic states push both processes at once, sometimes in opposite directions. As a result, a health claim can be accurate about one side and still mislead about the net outcome. "Captopril protects the kidneys in diabetes" is supported, yet the same study finds the drug speeds up bone loss by increasing breakdown and suppressing formation (Differential response of bone and kidney to ACEI in db/db mice). Current fact-checkers give such a claim a single verdict. Because they only check what the claim says, they cannot flag the harm it leaves out.
We propose to verify bone health claims by building a small, fixed "effect profile" from the evidence itself. The profile records the direction of effect on bone formation, bone resorption, net bone mass, and any other organ mentioned, together with the population the finding comes from (cell culture, ovariectomized mouse, aging mouse, human cohort). We then compare that profile with what the claim asserts. The comparison yields verdicts that a single scalar or a true/false label cannot express: supported, contradicted, misleading by omission (every stated part is true, but the evidence shows an unstated opposing effect on bone), and overextended (true in mice, asserted for people). The new element is the evidence-side profile. Existing decomposition methods split the claim into parts, so they can only check what the claim contains. Our method lays out what the evidence contains and asks what the claim left out.
Why bone remodeling is the right testbed
The misinformation literature agrees that real misinformation lives in gray areas such as cherry-picking, omitted context and true facts placed in misleading frames (A Survey on Automated Fact-Checking). Yet "misleading by omission" is hard to annotate in general, because almost anything could have been omitted. Bone remodeling makes the problem well posed. Bone papers routinely measure both sides: bone formation rate, osteoblast markers, counts of osteoclasts stained with the TRAP enzyme marker, and bone density. So there is a short, standard list of axes along which an omission can matter.
The biology in our file is full of exactly these two-sided findings. In the hepcidin-1 knockout mouse, iron overload both suppresses formation and promotes resorption (Iron accumulation induced by hepcidin1 knockout). Losing the protein Kindlin-2 lowers pro-formation Wnt signaling while raising pro-resorption RANKL signaling (Kindlin-2 regulates bone homeostasis). The inflammatory signal TNFα drives resorption while repressing formation, and the enzyme RSK2 opposes only the formation-side effect (RSK2 protects against inflammatory arthritis-induced bone destruction). Osteoclasts can suppress osteoblasts from a distance through small RNA carried in vesicles, as with miR-214-3p (Osteoclast-derived exosomal miR-214-3p). Closely related receptors can act in opposite directions, with PPARβ favoring bone formation and PPARγ disfavoring it (PPARβ regulates bone-metabolism). The enzyme USP7 brakes osteoclast formation through two separate routes (USP7 inhibits osteoclastogenesis). Many of these findings come only from mice or cell culture, which invites overextension to humans.
Bone biology therefore supplies both the structure of the evidence profile and a natural generator of realistic half-truths. Misinformation detection supplies the task. We claim nothing about bone biology itself.
Core Claim
We claim that verifying a claim against an evidence-derived profile over fixed axes will catch omission-type and population-overextension errors that claim-driven verifiers miss, without losing accuracy on ordinary supported/contradicted claims. The evidence-side extraction is the ingredient that matters. If a verifier given identical evidence and label definitions, but no explicit profile, catches omissions just as well, the claim is false.
Research Program
The system has four stages.
The first stage is retrieval. We retrieve full abstracts rather than search snippets, because snippets are known to be insufficient: they sufficed for only 45% of claims versus 83% for full pages in X-Fact, and snippet-based checking reached only 44.4 macro F1 on HealthFC.
The second stage builds the evidence profile. A language model fills a fixed schema from each retrieved abstract. For each intervention–outcome pair it records the direction of effect on formation, resorption, net bone and any other named organ (increase, decrease, no change, not measured), a pointer to the supporting sentence, and a population tag. This works like the evidence distillation in the coarse-to-fine cascaded network and the sentence localization in ReSCo-CC, but we extract slots rather than salient text. Each filled slot must quote a sentence, and we check the quote with an entailment model before accepting it.
The third stage maps the claim onto the same schema. Axes the claim does not address are marked "unstated," and the claim's implied net message about bone (good, bad, neutral) is recorded. Subtype and cell-type conflations, such as PPARβ versus PPARγ or an osteoclast effect stated as an osteoblast effect, are resolved at this step. This extends the disambiguation idea of Dis2Dis, which rewrote general linguistic ambiguity, to biological entity ambiguity.
The fourth stage derives the verdict from the two profiles with a small set of explicit rules. Any stated slot that the evidence contradicts gives contradicted. If all stated slots are supported but an unstated slot runs against the claim's implied net message, the verdict is misleading by omission. If a slot is supported only in a non-human population while the claim implies humans, the verdict is overextended. The explanation is written from the filled slots and their quoted sentences, so by construction it cannot contradict the label. That addresses the label–explanation mismatch documented on PUBHEALTH.
This design differs from the closest verifiers in specific ways. BiDeV uses role-played LLM agents to rewrite vague claims, split them into atomic parts and filter redundant evidence. It was tested on Wikipedia multi-hop claims, and its parts come from the claim, so an omitted effect never becomes a sub-question. DiVA pairs agentic search with a scalar factuality score, and a single scalar cannot say "true on one axis, harmful on another." CoVer weighs conflicting evidence by directness and reliability, which suits social-media noise. In bone biology, though, the apparent conflicts are usually both true, arising from different axes or populations. The age-dependent link between uric acid and bone density is an example (serum uric acid and lumbar BMD). Our population tag turns such conflicts into scoped findings instead of noise to be discarded.
Data
No bone-remodeling fact-checking benchmark exists in the material, so we will build one, called BoneBalance. We will collect roughly 1,500 abstracts on osteoporosis, osteoarthritis subchondral bone, inflammatory arthritis, drug-induced bone loss and iron or oxidative metabolism. The abstracts in our file serve as seed and pilot items.
Two annotators with bone-biology training will fill the gold evidence profile for each abstract. They will then write claims in five types: faithful, contradicted, omission, overextended and subtype-conflated. Omission claims are written only where the gold profile actually contains an opposing unstated effect.
Controlled claims written by annotators risk stylistic artifacts. Claim-only models already reach 38.2 macro F1 on X-Fact in-domain against 41.9 with evidence. We will therefore also collect a natural test set: bone- and osteoporosis-related claims filtered from PUBHEALTH and HealthFC, re-labeled with our scheme by the same annotators. We will report agreement (Cohen's kappa) per verdict type. If agreement on "misleading by omission" is low, the category itself is unsound, and we would learn that before building models.
Baselines
The naive reference point is a claim-only DeBERTa classifier fine-tuned on BoneBalance training claims. This model measures how much the claim wording alone reveals. On HealthFC, the claim-only model reached 39.3 macro F1 against 81.9 with oracle evidence (HealthFC). A method that does not clearly beat it on the natural test set is exploiting artifacts.
The strongest fine-tuned reader is a SciBERT veracity classifier over the same retrieved abstracts. It is trained on BoneBalance with our label set, following the recipe that gave 70.52 macro F1 on PUBHEALTH's four-way task (Explainable Automated Fact-Checking for Public Health Claims). We also include the joint evidence-and-verdict DeBERTa model that reported 67.5 macro F1 on HealthFC. These are the strongest supervised competitors because they already beat general-domain encoders on health claims.
Among LLM systems, the strongest competitor is direct prompting of the same backbone model our pipeline uses. The model receives the full retrieved abstracts, detailed definitions of all five verdicts and worked examples of each. It is our main threat, because a capable model might infer omissions without an explicit profile. We also run a BiDeV-style claim decomposition pipeline on the same backbone, to test claim-side decomposition directly.
All systems see the same retrieval results and the same backbone where applicable. Every LLM system gets the same budget of prompt revisions tuned on a shared development split, and every fine-tuned model gets the same hyperparameter search budget. We report oracle-abstract and retrieved-abstract settings separately.
Two ablations isolate the key ingredient. The first keeps the pipeline but replaces the fixed-axis evidence profile with a free-form evidence summary. The second keeps the axes but fills them from the claim only, dropping evidence-side profiling. Beating a baseline means a gain in F1 on the omission and overextended classes of at least 10 points on the natural test set, with 95% confidence intervals from paired bootstrap resampling, and no significant loss in overall macro F1.
Why not just…?
Why not just prompt a strong LLM with the abstract and good label definitions? This is a real threat, and it is the comparison the whole proposal is organized around. Prior evidence suggests unstructured generation struggles as reasoning gets more involved. A naive generative verifier scored 57.14 Precision@1 on Musique against 79.87 for DiVA's structured pipeline (DiVA). Few-shot Llama-3 also hallucinated ungrounded details more often than a small fine-tuned model in Dis2Dis. Neither result is about omissions, however. If the direct-prompting baseline matches us on the omission class, we will say so. The contribution would then shrink to the benchmark and the profile as an evaluation and explanation tool, not a better detector.
Why not just use claim decomposition like BiDeV? Decomposition produces sub-questions only about content the claim contains. A claim that mentions only kidney protection yields no sub-question about resorption. BiDeV's gains, 17.72% on 4-hop Hover claims (BiDeV), come from claims whose difficulty lies in what they state. The BiDeV-style baseline and the claim-only-axes ablation test this prediction directly. We expect both to do well on contradicted claims and poorly on omission claims.
Why not just fine-tune a biomedical classifier with a "mixture" label? PUBHEALTH already has a Mixture class, and SciBERT handles the four-way task reasonably well. A single "mixture" label, though, does not say which part is wrong or what is missing, and its explanations can contradict its label. The SciBERT baseline is trained on our finer labels, so the comparison is fair. We will also report its confusion between omission and supported, the error that matters for readers.
Proposed Experiments
The decisive experiment compares omission-class and overextended-class F1 on the natural test set across all systems, in both oracle and retrieved settings. The ablation replacing the profile with a free-form summary tells us whether the explicit axes, not just extra reasoning steps, carry the gain.
A second check measures profile extraction accuracy against gold profiles, slot by slot. This separates errors in reading the evidence from errors in comparing it with the claim. We expect the retrieved-evidence drop to be substantial, as on HealthFC, where veracity fell from 81.9 with oracle evidence to 65.1 in the pipeline system.
A third check measures label–explanation consistency with an entailment model and a blind expert rating of 200 explanations. This tests whether writing the explanation from the filled slots removes the mismatch reported on PUBHEALTH.
Finally, as a transfer probe, we run the unchanged pipeline on the full PUBHEALTH test set and report Mixture-class F1 against the SciBERT result. We make no claim that bone-specific axes generalize. We only ask whether they hurt performance outside bone.
References
- Differential response of bone and kidney to ACEI in db/db mice: A potential effect of captopril on accelerating bone loss
- A Survey on Automated Fact-Checking
- Iron accumulation induced by hepcidin1 knockout accelerates the progression of aging osteoporosis
- Focal adhesion protein Kindlin-2 regulates bone homeostasis in mice
- ERK-activated kinase RSK2 protects against inflammatory arthritis-induced bone destruction by opposing the tumour necrosis factor α-mediated inhibition of bone formation
- Osteoclast-derived exosomal miR-214-3p inhibits osteoblastic bone formation
- PPARβ regulates bone-metabolism by facilitating Wnt-signalling
- USP7 Inhibits Osteoclastogenesis via Dual Effects of Attenuating TRAF6/TAK1 Axis and Stimulating STING Signaling
- X-Fact: A New Benchmark Dataset for Multilingual Fact Checking
- HealthFC: Verifying Health Claims with Evidence-Based Medical Fact-Checking
- A Coarse-to-fine Cascaded Evidence-Distillation Neural Network for Explainable Fake News Detection
- ReSCo-CC: Unsupervised Identification of Key Disinformation Sentences
- Dis2Dis: Explaining Ambiguity in Fact-Checking
- Explainable Automated Fact-Checking for Public Health Claims
- BiDeV: Bilateral Defusing Verification for Complex Claim Fact-Checking
- DiVA: Fine-grained Factuality Verification with Agentic-Discriminative Verifier
- Age-stratified association between serum uric acid and lumbar bone mineral density in elderly Chinese women with vertebral compression fractures: a cross-sectional analysis