Abstract
Background
Depression is currently underdiagnosed among older adults. As part of the Novel Assessment of Nutrition and Aging (NANA) validation study, 40 older adults self-reported their mood using a touchscreen computer over three, one-week periods. Here, we demonstrate the potential of these data to predict future depression status.
Methods
We analysed data from the NANA validation study using a machine learning approach. We applied the least absolute shrinkage and selection operator with a logistic model to averages of six measures of mood, with depression status according to the Geriatric Depression Scale 10 weeks later as the outcome variable. We tested multiple values of the selection parameter in order to produce a model with low deviance. We used a cross-validation framework to avoid overspecialisation, and receiver operating characteristic curve (ROC) analysis to determine the quality of the fitted model.
Results
The model we report contained coefficients for two variables: sadness and tiredness, as well as a constant. The cross-validated area under the ROC curve for this model was 0.88 (CI: 0.69 to 0.97).
Limitations
While results are based on a small sample, the methodology for the selection of variables appears suitable for the problem at hand, suggesting promise for a wider study and ultimate deployment with older adults at increased risk of depression.
Conclusions
We have identified self-reported scales of sadness and tiredness as sensitive measures which have the potential to predict future depression status in older adults, partially addressing the problem of underdiagnosis.
Depression is currently underdiagnosed among older adults. As part of the Novel Assessment of Nutrition and Aging (NANA) validation study, 40 older adults self-reported their mood using a touchscreen computer over three, one-week periods. Here, we demonstrate the potential of these data to predict future depression status.
Methods
We analysed data from the NANA validation study using a machine learning approach. We applied the least absolute shrinkage and selection operator with a logistic model to averages of six measures of mood, with depression status according to the Geriatric Depression Scale 10 weeks later as the outcome variable. We tested multiple values of the selection parameter in order to produce a model with low deviance. We used a cross-validation framework to avoid overspecialisation, and receiver operating characteristic curve (ROC) analysis to determine the quality of the fitted model.
Results
The model we report contained coefficients for two variables: sadness and tiredness, as well as a constant. The cross-validated area under the ROC curve for this model was 0.88 (CI: 0.69 to 0.97).
Limitations
While results are based on a small sample, the methodology for the selection of variables appears suitable for the problem at hand, suggesting promise for a wider study and ultimate deployment with older adults at increased risk of depression.
Conclusions
We have identified self-reported scales of sadness and tiredness as sensitive measures which have the potential to predict future depression status in older adults, partially addressing the problem of underdiagnosis.
| Original language | English |
|---|---|
| Journal | Journal of Affective Disorders |
| Early online date | 14 Feb 2017 |
| DOIs | |
| Publication status | Published - 2017 |