# Liver Attenuation Task Card Task id: `liver_attenuation` Evidence status: grounded Evidence reviewed: 2026-07-28 Benchmark status: complete ## Target Hologic/SuperSonic MACH 30 Att.PLUS attenuation coefficient. It quantifies frequency-normalized ultrasound energy loss and generally rises with hepatic fat. Implementation facts: - target field: `c__attenuation` - target dataset: `liver_ultrasound` - target source: None (loaded through HPP `hpp_data_loader`) - default research stage: `00_00_visit` - valid range: 0.10-1.50 dB/cm/MHz; a PhenoBench plausibility filter, not a device or diagnostic interval - primary metric: R2 ## Why This Matters Liver attenuation adds a hepatic-fat imaging marker distinct from blood scores. Forecasting it does not establish MASLD, histologic grade, causality, or intervention response. ## Related HPP Measurements HPP work jointly analyzed gut-microbiome pathways with liver attenuation, CGM, and DXA. Same-visit comparators include sound speed, viscosity, elasticity, FLI, and liver blood tests; this Task does not establish their relationships. ## Benchmark-Track Evidence literature review prioritized a fixed Nightingale metabolic panel, then DXA regional body composition, for hepatic-fat prediction beyond age/sex/BMI. Metabolomic profiles show associations with liver fat distinct from visceral fat, while DXA android/gynoid and visceral fat capture adiposity not represented by BMI alone. HPP microbiome work linked many microbial features with liver attenuation, but that evidence was cross-sectional and did not justify prioritizing a microbiome track over Nightingale or DXA. ## V0 Benchmark Execution The frozen same-visit V1→V1 matrix completed on 2026-07-29. All six rows are locally successful, remotely `FINISHED` in operational source, and share the exact cohort within each track. - Nightingale: n=7,065; test=1,066; best R2=0.131 (Ridge), with paired delta R2 +0.047 [0.022, 0.071] over age/sex/BMI. - DXA: n=8,594; test=1,284; best R2=0.130 (Ridge), with paired delta R2 +0.053 [0.032, 0.072]. - Full rows and provenance: deep dive and aggregate review. ## Clinically Meaningful Variants Continuous attenuation remains primary. A same-platform, single-center NAFLD study reported Att.PLUS thresholds of 0.46/0.50/0.52 dB/cm/MHz for at least S1/S2/S3 (AUROC 0.82/0.70/0.73). These are hypothesis-only leads: even a binary at-least-S1 Task requires HPP acquisition/software, population, balance, and repeatability validation. Canon ATI and CAP thresholds are not interchangeable. Published thresholds vary across studies and are not used as benchmark labels. ## Baselines And Ceilings `NoFeaturesPredictor` contributes no modality features; `DemographicFloorStrategy` fits age, sex, and BMI. Last observation carried forward (LOCF) copies prior Y. Ordinary least squares (OLS) and gradient-boosted trees (GBDT) fit prior Y plus demographics. HealthFormer reported LOCF and per-modality linear baselines, not trees or the demographic floor. Compare representations on one cohort. ## Caveats - The dataset has no formal global quality flag; acquisition failures are filtered upstream. - Operator, protocol, body depth/BMI, and device algorithm affect values. - Strong autocorrelation can make LOCF difficult to beat. ## Metrics R2 is primary. Also report RMSE, MAE, Pearson r, Spearman rho, and same-cohort deltas against LOCF. Review age/BMI residuals and target-range coverage. ## References - Lutsker et al. (2026), HealthFormer, . - HPP `liver_ultrasound` dataset documentation for `c__attenuation`. - Tanpowpong et al. (2024), Att.PLUS versus MRI-PDFF, . - Keshet and Segal (2024), HPP microbiome and metabolic-health map, . - Pang et al. (2022), adiposity, metabolomics, and NAFLD risk, . - Lind et al. (2021), metabolomic profiles of liver and visceral fat, . ## Open Questions - Does HPP repeatability support a device-matched binary threshold? - **Which Att.PLUS grade boundaries are operative for HPP - 0.46 (Tanpowpong 2024, this card) or 0.63 (Popa 2021, the HPP dataset docs)?** Open at HPP documentation. Whichever wins, both repositories should name the same study, and this card needs the corresponding edit. Blocked on that reconciliation, deliberately not guessed here. - Where does the `0.31 dB/cm/MHz` "steatosis suspected" cutoff come from? It is unsourced in the HPP dataset docs yet drives their headline prevalence figure and ships as binning code. Same issue.