What each model estimates, how it is evaluated and what the results show, from predictive performance to calibration and classification performance.
Internal analytical evaluation3 regression models · 2 classifiersResearch use
Scientific status: software checks and primary data
Software corrections have been prepared and tested: fitting with preprocessing learned only on the training data, consistent metric calculation and checks of the original HSE fields. The models have not been retrained on the study data; no new individual predictions or external scientific validation are claimed.
HSE: the original formula defines a percentage scale. Text responses were found to have been converted to calendar dates, and 24 hours were substituted when times were missing. Range-based responses and impossible combinations require review against the codebook; there is no automatic division by 60, averaging or clipping.
HbA1c: models using these variables require traceable laboratory-field units and a documented target transformation. Categories in the manuscript do not confirm the units used by the separate HbA1c V2 model.
Estradiol: different censoring limits in the import require reconciliation of laboratory methods and rules for handling < and > signs. A sign is not replaced with a point value.
Historical exports and internal evaluations are retained below. A code correction does not retrospectively correct primary responses or establish the reliability of an individual result. Independent platform functions are preserved. Information about prepared code does not confirm that the live API has been updated.
6 720study observations
367prepared features
5research models
3analytical domains
01 /
Results in charts
Different tasks require different evaluation methods. Explained variation is shown for regression; classification performance gives equal weight to each class.
Regression models
R² · reproducing the target measure
The models have different targets. The high CoreClockScore R² describes reproduction of an age-associated index: age is included among the input features.Categorical models
Macro-F1 · saved holdout evaluation
Macro-F1 averages the F1 scores of the three classes with equal weight. Saved holdout results are shown here. The confusion matrices represent separate evaluation data; the differences are explicitly described below.
What these results mean
CoreClockScore describes an age-associated index. Continuous rTL and TelomereRiskScore reflect a small explained component of target variation. For classifiers, overall accuracy, performance within each class and a simple benchmark, the majority-class rule, should be considered together.
02 /
What the five models show
Each model's algorithm, task and output are presented alongside internal evaluation metrics. Values in the cards are rounded; regression targets use different scales, so MAE and RMSE should be compared within a task. Exported metrics and matrix-derived metrics are shown separately for classifiers.
Model 01RegressionLightGBM48 features in the model specification
CoreClockScore Beta
A composite age-associated index for research on age-related profiles.
Internal holdout evaluation
0.990R²
0.019RMSE
0.003MAE
Age is included among the features and has the highest importance estimate. The high R² relates to reproduction of the target index; an independent ageing signal requires separate analysis. RMSE and MAE are expressed on the index scale.
Target measure
CoreClockScore_beta
Input groups
Age, clinical and biochemical measures; imputed values and missingness indicators.
Model 02RegressionFastForest363 features in the model specification
rTL — continuous estimation
A continuous telomere-profile model for research comparisons.
Internal holdout evaluation
0.035R²
8.155RMSE
6.409MAE
R² ≈ 0.035 corresponds to about 3.5% explained variation. MAE ≈ 6.409 and RMSE ≈ 8.155 are points on the analytical rTL_0_100 scale (0–100). The construction of this scale must be distinguished from the laboratory T/S ratio.
Target measure
rTL_0_100
Input groups
Anthropometric, hormonal, clinical-epidemiological and questionnaire measures; missingness indicators.
Model 03ClassificationFastForest363 features in the model specification
rTL — categorical stratification
Three analytical telomere-profile classes: short, intermediate and long.
Saved holdout result
0.351Macro-F1
0.366MacroAccuracy · export
0.9803LogLoss
The export stored MacroAccuracy in the Accuracy field. This is not the proportion of all predictions that are correct. The saved-result metrics and the matrix recalculations below refer to different evaluation artefacts; they have not yet been reconciled to a single run.
Target measure
rTL_class
Input groups
Anthropometric, biochemical, clinical-epidemiological and questionnaire measures, including missingness indicators for hormonal and reproductive features.
Model 04RegressionFastForest39 features in the model specification
TelomereRiskScore — continuous estimation
A continuous telomere-associated index of the age-metabolic profile.
Internal holdout evaluation
0.044R²
0.202RMSE
0.158MAE
R² ≈ 0.044 corresponds to about 4.4% explained variation. Errors are expressed on the TelomereRiskScore scale, a calculated index rather than the probability of a clinical event.
Target measure
TelomereRiskScore
Input groups
Anthropometry, sex, age, education, sleep and behavioural measures.
Model 05ClassificationFastForest39 features in the model specification
TelomereRiskScore — categorical stratification
Three analytical TelomereRiskScore levels: low, intermediate and high.
Saved holdout result
0.389Macro-F1
0.391MacroAccuracy · export
1.1135LogLoss
Target-index levels: below −0.1, from −0.1 to 0.1 and above 0.1. Metrics from the detailed holdout block are shown; recalculation of the saved matrix is considered separately.
Target measure
TRS_Class
Input groups
Waist circumference, sex, age, BMI, sleep measures, education features and missingness indicators.
R² is explained variation relative to the mean; MAE is mean absolute error; RMSE is root mean squared error. Standard Accuracy is the proportion of correctly classified observations; MacroAccuracy is mean class recall (previously mislabelled Accuracy in the export); Macro-F1 is mean class F1; LogLoss evaluates probabilistic predictions, with lower values indicating better performance.
03 /
Calibration profiles
Each point represents the means for one of ten calibration groups. Proximity to the diagonal indicates agreement between the mean prediction and the mean observed value. Model scales differ.
Group meanEquality of prediction and observation
CoreClockScore Beta
10 groups · total n = 1 343. This model's target scale.
Values by group
Group
n
Prediction
Observed
1
135
0.212
0.218
2
134
0.283
0.286
3
134
0.339
0.340
4
134
0.393
0.397
5
135
0.448
0.449
6
134
0.511
0.512
7
134
0.594
0.591
8
134
0.668
0.666
9
134
0.742
0.741
10
135
0.829
0.824
rTL · continuous model
10 groups · total n = 1 344. This model's target scale.
Values by group
Group
n
Prediction
Observed
1
135
28.414
38.211
2
134
38.113
44.063
3
134
42.501
47.227
4
135
46.279
48.786
5
134
49.207
49.846
6
134
51.856
51.228
7
135
54.512
52.976
8
134
57.668
54.196
9
134
61.435
56.252
10
135
68.686
58.141
TelomereRiskScore · continuous model
10 groups · total n = 1 343. This model's target scale.
Values by group
Group
n
Prediction
Observed
1
135
-0.503
-0.211
2
134
-0.301
-0.149
3
134
-0.202
-0.135
4
134
-0.123
-0.072
5
135
-0.057
-0.043
6
134
0.007
-0.006
7
134
0.082
0.033
8
134
0.174
0.096
9
134
0.292
0.131
10
135
0.524
0.305
The charts compare group means and show aggregate agreement between predictions and observations. The distribution of individual errors requires separate analysis.
04 /
What the classifiers show
Two sets of evaluation data
Source-code check: ClassificationTrainer.cs, line 220, writes MacroAccuracy to a field named Accuracy. Thus, 0.366433 is the macro-averaged export metric, not standard accuracy. From the saved rTL matrix, standard Accuracy = 762/1344 = 0.566964, while mean class recall ≈ 0.369696. The latter also differs from 0.366433: the source code uses separate training runs (lines 199, 365 and 134), so the artefacts have not yet been reconciled to a single run. Values are retained unchanged; this is a correction of interpretation, not retraining.
Model
Evidence source
Metric from that source
Macro-F1
LogLoss
rTL
Saved holdout · MacroAccuracy (Accuracy field)
0.366433
0.350991
0.980289
rTL
Recalculated matrix · standard Accuracy · n = 1 344
0.566964
0.353547
—
TelomereRiskScore
Saved holdout · MacroAccuracy (Accuracy field)
0.390600
0.389100
1.113500
TelomereRiskScore
Recalculated matrix · standard Accuracy · n = 1 343
0.390916
0.389128
—
All matrices, recall values and majority-rule comparisons below refer specifically to matrix recalculations. Regression fields incorrectly included in the packaged TRS classifier metadata are not used as classification metrics.
Matrices and their recalculated metrics
The matrices show correct predictions and errors by class. Accuracy, Macro-F1, class recall and the majority-class rule are calculated directly from these values.
rTL · categorical model
n = 1 344 · rows: true class; columns: prediction
True ↓ Predicted →
Short
Intermediate
Long
Short
24
227
28
Intermediate
24
697
85
Long
4
214
41
Recall: short — 8.6%; intermediate — 86.5%; long — 15.8%. Matrix-derived Macro-F1: 0.354.
A common 0–100% scale. The majority-class rule is calculated from the true classes in each matrix.
Interpreting the comparison
56,7%
rTL: The model's Accuracy is below the 60.0% achieved by the majority-class rule. Recall for the extreme classes is 8.6% and 15.8%; most predictions fall in the intermediate class.
39,1%
TelomereRiskScore: The model's Accuracy exceeds the 37.5% achieved by the majority-class rule by 1.6 percentage points. Recall for the three classes is approximately 38–40%.
This is a descriptive comparison based on matrix data. The statistical significance of the difference and generalisability to other samples require separate evaluation.
How the metrics are calculated
Accuracy = sum of the diagonal / n. Class recall = correct predictions for the class / number of true observations in the class. Class F1 = 2 × correct predictions / (number of true + number of predicted observations in the class). Macro-F1 is the mean of the three F1 values. Majority-rule Accuracy = size of the largest true class / n.
05 /
How internal evaluation was performed
The described workflow includes feature preparation, a held-out evaluation sample and algorithm selection using a prespecified criterion.
STEP 1
Preparation
Median imputation, missingness indicators and exclusion lists. A total of 247 feature rules were retained; automatic outlier processing was disabled.
Minimise MAE for regression; maximise Macro-F1 for classification.
STEP 4
Result checks
Holdout metrics, calibration means, confusion matrices and checks of evaluation artefacts.
Data level
Count
Meaning
Original cohort
6 720
The complete original set of observations.
Prepared ModelTrainer dataset
6 718
5 374 train + 1 344 test.
Source columns
261
Structure of the original ModelTrainer dataset.
Prepared features
367
The processed set; subsets differ between models.
Input-schema parameters
238
The input schema is a separate representation, not a replacement for the number of prepared features.
Evaluation status and interpretation
Internal holdout evaluation describes performance on held-out data from this research dataset. Independent external and clinical evaluation of these five models is not presented.
The configuration contains KFold = 5, but the detailed results are labelled Holdout: they should be interpreted as held-out evaluation, not the average of five folds.
Exclusion lists and the blacklist are part of feature control; their presence alone does not establish the complete absence of leakage. For CoreClock, a separate analysis excluding chronological age is particularly important. ModelTrainer distinguishes CoreClockScore_beta and CoreClockScore_pct targets; the beta model specification is presented here, without conflating the two configurations.
The stated bootstrap uses 1 000 iterations; the composition of resamples and intervals for the main metrics require a separate protocol. No individual confidence intervals are claimed here.
Each calibration table and matrix retains its own sample size: n = 1 343 or 1 344. These totals are not replaced with the overall test-set size.
06 /
Why these algorithms were selected
Comparing the saved candidates helps explain the selection criterion. LightGBM has a lower MAE for CoreClock, although FastTree has a slightly higher R². The classifier comparisons below use saved holdout results, not matrix recalculations.
CoreClockScore Beta
Selection criterion: minimum MAE.
Algorithm
R²
RMSE
MAE
FastForest
0.911425
0.0573265
0.0455986
LbfgsPoisson
0.855367
0.0732545
0.0460476
LightGBM · selected
0.990275
0.0189957
0.00274859
FastTree
0.990599
0.0186757
0.00286867
Sdca
0.957268
0.0398179
0.0246531
rTL — continuous estimation
Selection criterion: minimum MAE.
Algorithm
R²
RMSE
MAE
FastForest · selected
0.0349046
8.1549
6.40946
LbfgsPoisson
-1.45330e+56
1.00072e+29
2.72968e+27
LightGBM
-0.0375755
8.45558
6.58218
FastTree
-0.0581061
8.53882
6.69325
Sdca
-4.87536e+06
18328.9
10715.4
rTL — categorical stratification
Selection criterion: maximum Macro-F1.
Algorithm
Macro-F1
MacroAccuracy · export
LogLoss
LightGBM
0.331752
0.361417
1.00788
FastForest · selected
0.350991
0.366433
0.980289
TelomereRiskScore — continuous estimation
Selection criterion: minimum MAE.
Algorithm
R²
RMSE
MAE
FastForest · selected
0.0437397
0.202298
0.15845
LightGBM
-0.0262205
0.209568
0.163176
FastTree
-0.0238055
0.209321
0.163329
Sdca
-0.0530215
0.212287
0.16579
TelomereRiskScore — categorical stratification
Selection criterion: maximum Macro-F1.
Algorithm
Macro-F1
MacroAccuracy · export
LogLoss
LightGBM
0.3692
0.3706
1.151
FastForest · selected
0.3891
0.3906
1.1135
For rTL, the LbfgsPoisson and Sdca candidates produced numerically unstable results; their values are retained, with scientific notation used for extreme values. Errors from models with different target scales must not be compared directly.
07 /
Data quality and feature preparation
Quality control covers completeness, ranges and numeric formats. The table below uses the prepared ModelTrainer dataset: n = 6 718.
Measure
Type of check
Count
Review
PSQ5i
Missing after cleaning
6 718
Check the variable's availability and role.
HSE
Missing after cleaning
6 697
Check coding and permissible ranges.
HSE
Out of range
6 622
Compare range rules with the original data dictionary.
Estradiol
Numeric parsing errors
1 534
Check laboratory-value formats.
Estradiol
Missing after cleaning
4 798
Clarify the pattern of availability and the imputation rules.
Thigh circumference · Thigh
Out of range
562
Check units and source records.
PSQ5i is missing after cleaning in 100% of rows, and HSE in approximately 99.69%. Missingness, out-of-range values and parsing errors may overlap: these counts must not be added to obtain a participant count. Reasons for missing values and their effect on a given model require analysis of the preparation process.
08 /
Molecular measurements: a separate stream
MSP panel: 478 laboratory samples
The laboratory molecular panel and calculated models are presented separately.
rTL_0_100 is an analytical 0–100 scale. Classes and indices are used for research comparisons.