An article on the results of long-term research at the Harvard Medical School was published in Nature Communications in the public domain (also available as a preprint on BioRxiv). First, the authors monitored the health of 60 laboratory mice until their natural death. For over a year, various health and fitness metrics were measured in mice, such as loss of visual and hearing acuity, gait, curvature of the back, etc. This training dataset was used to develop two machine learning models. One of them, on the basis of "aging indicators", determines the biological age of the mouse; the other is used to predict the remaining lifespan of the organism. Further observations showed that the models were accurate within two months.
Frailty Indices (FI) are some complex indicators of health deterioration with age. They are non-invasive, that is, they do not affect the object and do not require excessive resources to use. There are adapted methods for determining the physical condition and calculating indicators of aging for mice. However, it is not known how accurately these or those of their varieties determine the state of the organism as a whole and predict the time of its life. The FI coefficient for mice is calculated out of 31 health scores, each of which is assigned a value of 0, 0.5 and 1 (1 indicates the corresponding age-related disability, 0 indicates its absence). These indicators include, for example, paw grip strength, tail elasticity, weakening of vision/hearing compared to a young mouse, etc. Biologists interested or working in this field can learn more about this technique from the video from the article authors at the end of this post.
The used machine learning technique is based on decision trees, structures in the form of tree (branching) graphs, at the nodes of which there are individual parameters that make up the FI index. Such trees are created from a training dataset, that is, all measured parameters for mice and real data on their lifespan, and then allow predictions of lifespan for new individuals based on their health indicators. The authors applied an improved decision tree learning algorithm called a random forest, or ensemble decision tree classifier. Its essence lies in the fact that not one decision tree is used, which on a small sample may not be completely accurate, but several (large number) of such trees, each of which is built on a random sample of the original data set. In this case, the predictive model is built in the form of averaging over a set of such trees (committee of decision trees), which improves the quality of predictions. Thus, the algorithm used 1000 such decision trees based on the same number of random samples of training data.
The main intended application of the developed model is to evaluate the effect of various clinical and therapeutic interventions presumably increasing the lifespan, for example, diet or any drug. Longitudinal studies, in this case, would have to last about three years (average lifetime of a mouse) for any drug. Therefore, such predictive biometrics is a tool that allows reducing the research time, having received at least preliminary conclusions about how promising the chosen direction is.
At the next stage, predictive models were tested on groups of mice, in which the effect of life-prolonging enzymes had been previously tested with measurement of the FI index. The researchers state that the proposed machine learning model is able to correctly predict the efficacy of appropriate therapy for increasing life expectancy.
The AI system also led to the conclusion that certain aging indicators are more correlated with mouse's health status and life expectancy prospects in the future. For example, body tremors and a degree of hearing loss were found to be more related to biological age than vision and hair loss. It should be emphasized once again that these findings apply only to laboratory mice. So far, artificial intelligence is not able to predict human life expectancy. Significantly more decisive factors come into play here than for mice, and they turn out to be much more interconnected.
Similar indicators of FI aging exist for humans. Moreover, the FI method for mice is an adaptation of the indicators originally developed for humans. Unfortunately, the biologists do not have a reliable sample of data with such a systematic health monitoring of people aged 60 to 90 years, including data on mortality. However, in the future, they expect to develop a similar machine learning system for rapid assessment of life expectancy and effectiveness of various therapeutic interventions for life extension.
The authors of the work propose the researchers working in this area and experimenting with mice, to use their tool on a special website-calculator of FI indices and the corresponding indicators of biological age and life expectancy of a mouse predicted by AI models.
Photos are from open sources.