Sync linage2

README.md

@@ -10,7 +10,7 @@ tags:
 
 # linage2
 
-Principal-component clinical clock trained on 20-year mortality in the NHANES IV 1999-2000 wave and tested in the 2001-2002 wave. The 59 model features are log transformed where the reference specifies, robustly z-scored by sex against a healthy 40-50 year old NHANES reference, and folded at 6 MAD-scaled units; the folded vector is projected onto sex-specific singular vectors and scored by a sex-specific Cox model, so male and female samples run through entirely separate parameter sets. Chronological age is supplied in years and enters the Cox terms in months, where it is a genuine covariate rather than a cancelling offset. Sex is coded female = 1 and male = 0; a dataset with no female column scores every sample with the male parameters. C-reactive protein is supplied raw in mg/dL and takes a plain natural log with no floor, so a reading below detection coded as 0 folds to the -6 cap instead of being clamped the way the BioAge clocks clamp it. The input contract is wider than the 59 model features: total cholesterol, HDL cholesterol and triglycerides are consumed only by the Friedewald LDL and are not features themselves, and 26 questionnaire codes feed the comorbidity, self-reported-health and healthcare-use indices. An absent questionnaire block is the main hazard and it biases the estimate downward. The substituted values are the reference cohort's median profile everywhere except the comorbidity index: answering no to all 22 conditions gives 0 where the cohort's median is 1/22, so that one feature is substituted marginally healthier than the median. It costs almost nothing, because the index's entire 0 to 1 range moves the estimate by only 0.06 years. The self-reported-health index is what does the damage: its substitute of good, unchanged health is exactly the cohort median, and a subject who would have reported poor and worsening health reads about 5.3 years younger than they should. The healthcare-use index pushes the other way, by about 0.5 years for a subject who made 16 or more visits. An absent lipid panel substitutes a total cholesterol chosen so that the derived LDL lands on the reference median, avoiding the reference implementation's hard 0 mmol/L substitution; a NaN inside a lipid column that is present still takes that hard 0 path and lowers the estimate by roughly 0.35 years. The range check cannot see this at all: it resolves ranges over the clock's declared input features, and the comorbidity, self-reported-health and healthcare-use indices are derived inside the model rather than supplied, so no range check ever applies to them. Heed the missing-feature warning the prediction pipeline emits; for this clock it is the only signal that the questionnaire block was absent.
+Principal-component clinical clock trained on 20-year mortality in the NHANES IV 1999-2000 wave and tested in the 2001-2002 wave. The 59 model features are log transformed where the reference specifies, robustly z-scored by sex against a healthy 40-50 year old NHANES reference, and folded at 6 MAD-scaled units; the folded vector is projected onto sex-specific singular vectors and scored by a sex-specific Cox model, so male and female samples run through entirely separate parameter sets. Chronological age is supplied in years and enters the Cox terms in months, where it is a genuine covariate rather than a cancelling offset. Sex is coded female = 1 and male = 0; a dataset with no female column scores every sample with the male parameters. C-reactive protein is supplied raw in mg/dL and takes a plain natural log with no floor, so a reading below detection coded as 0 folds to the -6 cap instead of being clamped the way the BioAge clocks clamp it. The input contract is wider than the 59 model features: total cholesterol, HDL cholesterol and triglycerides are consumed only by the Friedewald LDL and are not features themselves, and 26 questionnaire codes feed the comorbidity, self-reported-health and healthcare-use indices. An absent questionnaire block is the main hazard and it biases the estimate downward. The substituted values are the reference cohort's median profile everywhere except the comorbidity index: answering no to all 22 conditions gives 0 where the cohort's median is 1/22, so that one feature is substituted marginally healthier than the median. It costs almost nothing, because the index's entire 0 to 1 range moves the estimate by only 0.06 years. The self-reported-health index is what does the damage: its substitute of good, unchanged health is exactly the cohort median, and a subject who would have reported poor and worsening health reads about 5.3 years younger than they should. The healthcare-use index is the raw NHANES 1999-2000 HUQ050 code (0 = none, 1 = 1, 2 = 2 to 3, 3 = 4 to 9, 4 = 10 to 12, 5 = 13 or more visits), so its valid range is 0 to 5 and the fold at 6 never fires on it; the 0 to 8 HUQ051 scale of later NHANES cycles must be collapsed onto it first (3 to 5 -> 3, 6 -> 4, 7 and 8 -> 5) or healthcare use is overstated. Substituting its median code of 2 for a subject who made 13 or more visits moves the male estimate up by about 0.35 years and the female estimate down by about 1.9 years. An absent lipid panel substitutes a total cholesterol chosen so that the derived LDL lands on the reference median, avoiding the reference implementation's hard 0 mmol/L substitution; a NaN inside a lipid column that is present still takes that hard 0 path and lowers the estimate by roughly 0.35 years. The range check cannot see this at all: it resolves ranges over the clock's declared input features, and the comorbidity, self-reported-health and healthcare-use indices are derived inside the model rather than supplied, so no range check ever applies to them. Heed the missing-feature warning the prediction pipeline emits; for this clock it is the only signal that the questionnaire block was absent.
 
 | | |
 |---|---|

config.json

@@ -10,7 +10,7 @@
  "last_author": "Jan Gruber",
  "model_type": "PCA + Cox regression",
  "n_features": 85,
- "notes": "Principal-component clinical clock trained on 20-year mortality in the NHANES IV 1999-2000 wave and tested in the 2001-2002 wave. The 59 model features are log transformed where the reference specifies, robustly z-scored by sex against a healthy 40-50 year old NHANES reference, and folded at 6 MAD-scaled units; the folded vector is projected onto sex-specific singular vectors and scored by a sex-specific Cox model, so male and female samples run through entirely separate parameter sets. Chronological age is supplied in years and enters the Cox terms in months, where it is a genuine covariate rather than a cancelling offset. Sex is coded female = 1 and male = 0; a dataset with no female column scores every sample with the male parameters. C-reactive protein is supplied raw in mg/dL and takes a plain natural log with no floor, so a reading below detection coded as 0 folds to the -6 cap instead of being clamped the way the BioAge clocks clamp it. The input contract is wider than the 59 model features: total cholesterol, HDL cholesterol and triglycerides are consumed only by the Friedewald LDL and are not features themselves, and 26 questionnaire codes feed the comorbidity, self-reported-health and healthcare-use indices. An absent questionnaire block is the main hazard and it biases the estimate downward. The substituted values are the reference cohort's median profile everywhere except the comorbidity index: answering no to all 22 conditions gives 0 where the cohort's median is 1/22, so that one feature is substituted marginally healthier than the median. It costs almost nothing, because the index's entire 0 to 1 range moves the estimate by only 0.06 years. The self-reported-health index is what does the damage: its substitute of good, unchanged health is exactly the cohort median, and a subject who would have reported poor and worsening health reads about 5.3 years younger than they should. The healthcare-use index pushes the other way, by about 0.5 years for a subject who made 16 or more visits. An absent lipid panel substitutes a total cholesterol chosen so that the derived LDL lands on the reference median, avoiding the reference implementation's hard 0 mmol/L substitution; a NaN inside a lipid column that is present still takes that hard 0 path and lowers the estimate by roughly 0.35 years. The range check cannot see this at all: it resolves ranges over the clock's declared input features, and the comorbidity, self-reported-health and healthcare-use indices are derived inside the model rather than supplied, so no range check ever applies to them. Heed the missing-feature warning the prediction pipeline emits; for this clock it is the only signal that the questionnaire block was absent.",
+ "notes": "Principal-component clinical clock trained on 20-year mortality in the NHANES IV 1999-2000 wave and tested in the 2001-2002 wave. The 59 model features are log transformed where the reference specifies, robustly z-scored by sex against a healthy 40-50 year old NHANES reference, and folded at 6 MAD-scaled units; the folded vector is projected onto sex-specific singular vectors and scored by a sex-specific Cox model, so male and female samples run through entirely separate parameter sets. Chronological age is supplied in years and enters the Cox terms in months, where it is a genuine covariate rather than a cancelling offset. Sex is coded female = 1 and male = 0; a dataset with no female column scores every sample with the male parameters. C-reactive protein is supplied raw in mg/dL and takes a plain natural log with no floor, so a reading below detection coded as 0 folds to the -6 cap instead of being clamped the way the BioAge clocks clamp it. The input contract is wider than the 59 model features: total cholesterol, HDL cholesterol and triglycerides are consumed only by the Friedewald LDL and are not features themselves, and 26 questionnaire codes feed the comorbidity, self-reported-health and healthcare-use indices. An absent questionnaire block is the main hazard and it biases the estimate downward. The substituted values are the reference cohort's median profile everywhere except the comorbidity index: answering no to all 22 conditions gives 0 where the cohort's median is 1/22, so that one feature is substituted marginally healthier than the median. It costs almost nothing, because the index's entire 0 to 1 range moves the estimate by only 0.06 years. The self-reported-health index is what does the damage: its substitute of good, unchanged health is exactly the cohort median, and a subject who would have reported poor and worsening health reads about 5.3 years younger than they should. The healthcare-use index is the raw NHANES 1999-2000 HUQ050 code (0 = none, 1 = 1, 2 = 2 to 3, 3 = 4 to 9, 4 = 10 to 12, 5 = 13 or more visits), so its valid range is 0 to 5 and the fold at 6 never fires on it; the 0 to 8 HUQ051 scale of later NHANES cycles must be collapsed onto it first (3 to 5 -> 3, 6 -> 4, 7 and 8 -> 5) or healthcare use is overstated. Substituting its median code of 2 for a subject who made 13 or more visits moves the male estimate up by about 0.35 years and the female estimate down by about 1.9 years. An absent lipid panel substitutes a total cholesterol chosen so that the derived LDL lands on the reference median, avoiding the reference implementation's hard 0 mmol/L substitution; a NaN inside a lipid column that is present still takes that hard 0 path and lowers the estimate by roughly 0.35 years. The range check cannot see this at all: it resolves ranges over the clock's declared input features, and the comorbidity, self-reported-health and healthcare-use indices are derived inside the model rather than supplied, so no range check ever applies to them. Heed the missing-feature warning the prediction pipeline emits; for this clock it is the only signal that the questionnaire block was absent.",
  "platform": [
   "clinical laboratory assays"
  ],

linage2.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b2726be7ff7d71d6c15b91d8598b639529c136252fdfa101d929189306168e28
-size 76263
+oid sha256:b342d21624c8f94c7468c23c6a1b214ed005fd47a267f22ce6b8fcf1a189651e
+size 76839