Supplementary Figures 1-5 Legends

Extended Data Fig. 1 Estimation of the scorer effect in clinical FI items.
A, The effect of tester varies across FI items. B, The estimated random effect across 4 scorers in the data set.


Extended Data Fig. 2 Detailed modeling analysis.
A, The distribution of age across 643 data points (533 mice). The distribution of manual FIadj scores across 643 data points (533 mice). B, To determine the contributions of frailty parameters in predicting Age, we calculated the feature importance of all frailty parameters. We discover that gait disorders, kyphosis and piloerection have the highest contributions. C, The random forest regression model performed better than other models with the lowest root-mean-squared error (RMSE) (n?=?50 independent train-test splits, p?<?2.2e???16, F3,147?=?59.53) and highest R2 (p?<?2.2e???16, F3,147?=?58.14) when compared using repeated-measures ANOVA. D, The vFRIGHT model performed better than the FRIGHT model with a lower RMSE (n?=?50 independent train-test splits, RMSEvFRIGHT?=?17.97?±?1.44, RMSEFRIGHT?=?20.62?±?4.78, p?<?6.1e???7, F1,49?=?32.84) and higher R2 (RMSEvFRIGHT =0.78?±?0.04, RMSEFRIGHT?=?0.76?±?0.07, p?<?2.1e???8, F1,49?=?44.54) when compared using repeated-measures ANOVA. E, The random forest regression model for predicting FI score on unseen future data performed better than all other models, with a lowest root-mean-squared error (RMSE) (n?=?50 independent train-test splits, p?<?8.3e???14, F3,147?=?26.62) and highest R2 (p?<?4.7e???14, F3,147?=?27.2). F, The plot shows the counts distribution (0 - green, 0.5 - orange, 1 - purple) for individual frailty parametersŃ for many parameters such as Nasal discharge, Rectal prolapse, Vaginal uterine and Diarrhea, the proportion of 0 counts is 1 (p0?=?1). Similarly, Dermatitis, Cataracts, Eye discharge swelling, Microphthalmia, Corneal opacity, Tail stiffening and Malocclusions have p0?>?0.95. G, The residuals versus the index and predicted versus true for training (Column 1; residual standard error = 8.5, difference in slopes (black vs gray) = 0.11) and test sets (Column 2; residual standard error = 15.87, difference in slopes (black vs gray) = 0.30) for the model that predicts Age using frailty index items for both training and test data. H, I, Out-of-bag (OOB) error based 95% prediction intervals (PIs) (gray lines) quantifying uncertainty in point estimates/predictions (gray dots). There is one interval per test mouse (n?=?107 unique mice, the test data contains some repeats of the same mice tested at different ages) and approximately 95% of the PI intervals contain the correct Age (red dots) and FI scores (blue dots). We ordered the x-axis (Test set index) in ascending order (from left to right) of the actual age/FI. The average PI width for all test mouseŐs predicted FI score is 0.18?±?0.04 (resp. 71.96?±?18.52 for the predicted Age), while the PI lengths range from 0.08 to 0.29 (resp. 28 to 113 for Age). n (C, D and E), the lower and upper hinges correspond to the first and third quartiles (the 25th and 75th percentiles) respectively, the line in the middle corresponds to the median, the upper (lower) hinge extends from the upper (lower) hinge to the largest (smallest) value not bigger (smaller) than 1.5 ? IQR where IQR is the interquartile range.


Extended Data Fig. 3 Correlation between video metrics.
A, Correlation between average/mean (x-axis) and median (y-axis) video gait metrics. The diagonal line corresponds to maximum correlation i.e. 1. B, Correlation between inter-quartile range (IQR, x-axis) and standard deviation (Stdev, y-axis) video gait metrics. The diagonal line corresponds to maximum correlation i.e. 1. A tight wrap of points around the diagonal line indicates a high correlation between mean and median or IQR and standard deviation for the respective metric.


Extended Data Fig. 4 Test for SimpsonŐs paradox.
A, Simpson (1951) showed that the statistical relationship observed in the population could be reversed within all of the subgroups that make up that population, leading to erroneous conclusions drawn from the population data. To test for the manifestation of SimpsonŐs paradox in our data, we split the bimodal Age distribution into two separate unimodal distributions (clusters), that is, less than 70 weeks old (L70, red) versus more than 70 weeks old (U70, blue). Next, we plotted the dependent variable (frailty) against each of the independent variables/features in our data and fit a simple linear regression model to each subgroup separately (solid red and blue lines) as well as to the aggregate data (black dotted line). B, We quantified the correlations by measuring the slope of the linear fits of the features (Y) on Age (X). We computed the slopes for L70, U70 and overall (All), then plotted the slopes for features in decreasing order of their relevance to the model (where we predict Age from these features). We went further and performed one-way ANOVA to test for differences in slopes between L70 and U70 sub-groups and the overall data (one-way ANOVA, F2,141?=?1.162, p?>?0.32). Next, we performed a false discovery rate adjusted post hoc pairwise comparisons using the t-test. We found no significant differences in the comparisons (L70 versus U70, p?=?0.38, L70 versus All, p?=?0.77 and U70 versus All, p?=?0.38). We found that SimpsonŐs paradox does not manifest in any of the top fifteen features in our data.


Extended Data Fig. 5 Further experiments to test model performance and parameters.
A, We compare the performance of different feature sets, 1) age alone, 2) video and 3) age + video, in predicting frailty across n?=?50 independent train-test splits. We use age alone as a feature in a linear (AgeL) and a generalized additive non-linear model (AgeG). Although we didnŐt notice a clear improvement of the random forest model (VideoRF) using video features over a vFI prediction based on age alone, a clear improvement in prediction performance is seen for the model (AllRF), which contains video features + age with lowest MSE (p?<?2.2e???16, F3,147?=?213.79, LMM post hoc pairwise comparison with AgeG, t147?=??12.21, FDR-adjusted p?<?.0001), lowest RMSE (p?<?2.2e???16, F3,147?=?172.88, LMM post hoc pairwise comparison with AgeG, t147?=??14.12, FDR-adjusted p?<?.0001) and highest R2 (p?<?2.2e???16, F3,147?=?171.12, LMM post hoc pairwise comparison with AgeG, t147?=?14.07, FDR-adjusted p?<?.0001). This shows that video features add important information pertaining to frailty that age alone does not. B, We picked animals whose ages and FI scores had an inverse relationship, that is, younger animals with higher FI scores and older animals with lower FI scores. We formed 5 test sets (n?=?43, 38, 45, 42, 20) containing animals with these criteria and trained the random forest (RF) model on the remaining mice. The model using only video features (VideoRF) does better than all other models for these mice with lowest MSE (p?<?1.6e???08, F3,12?=?91.07, LMM post hoc pairwise comparison with AgeG, t12?=?13.60, FDR-adjusted p?<?.0001), lowest RMSE (p?<?1.6e???08, F3,12?=?93.88, LMM post hoc pairwise comparison with AgeG, t12?=?14.15, FDR-adjusted p?<?.0001) and highest R2 (p?<?1.31e???08, F3,12?=?94.32, LMM post hoc pairwise comparison with AgeG, t12?=?14.10, FDR-adjusted p?<?.0001). C, We further investigate the difference between Age and vFI predictors in terms of feature importance. Features lying along the diagonal are important for both Age and vFI predictions. D, Predicting FI score from video features extracted from videos of shorter durations. We used video features generated from videos with shorter durations (first 5 and 20?minutes) to investigate the loss in accuracy in predicting age and FI score. We used the random forest model trained with features generated from 60-minute videos as a baseline model for comparison. We found a diminished loss in accuracy using shorter videos. The features associated with 60-minute videos had the best accuracy for vFI prediction (LMM where ÔsimulationŐ is the random effect, nsim = 50; lowest MAE, F2,98?=?178.39, p?<?2.2e???16; lowest RMSE, F2,98?=?156.93, p?<?2.2e???16); highest R2 (p?<?2.2e???16, F2,98?=?297.3). We observed a significant drop in performance accuracy when the open field test length is reduced from 60 to 20-minute video (LMM with post hoc pairwise comparisons - MAE, t98?=?14.82, FDR-adjusted p?<?0.0001; RMSE, t98?=?13.69, FDR-adjusted p?<?0.0001; R2, t98?=??19.22, FDR-adjusted p?<?0.0001). E, To see how much training data is realistically needed, we performed a simulation study (n?=?50 independent train-test splits) where we allocated different percentage of total data to training. As expected, there is a general downward (upward) trend in MAE, RMSE (R2 with an increasing percentage of data allocated to training set. Indeed, a smaller training set (< 80% training) can reach a similar training performance. In A, B, E and D, the lower and upper hinges correspond to the first and third quartiles (the 25th and 75th percentiles) respectively, the line in the middle corresponds to the median, the upper (lower) hinge extends from the upper (lower) hinge to the largest (smallest) value not bigger (smaller than 1.5 ? IQR where IQR is the interquartile range.