Applications of mathematics in clinical setting

Machine learning is on its way to provide the type of personalized health care referred to by the OP. In June of this year the FDA (US Food and Drug Administration) has proposed a regulatory framework for machine learning algorithms in medical decision making.

The document distinguishes "locked algorithms" (basically decision trees) and "adaptive algorithms" (a supervised learning algorithm that will "change its behavior using a defined learning process", as the inputs are given new data). Only the locked algorithms are currently approved by the FDA, and the document tries to set up a pathway for supervised machine learning algorithms to be used in health care.

An example of a simple mathematical/evolutionary game theory model used to determine treatment scheduling in clinical treatment of metastic and castrate resistant prostate cancer can be found at https://www.nature.com/articles/s41467-017-01968-5. While the clinical trial is on-going, initial results show that the model derived treatment schedule offers significant improvement in time to treatment failure when compared to the standard-of-care.

In short, the authors use a 3 population evolutionary game theory model to study the interplay between androgen dependent, androgen independent and androgen producing cancerous cells. They use this model to determine an "adaptive therapy" dosing strategy for abiraterone, a drug that strongly effects androgen dependent cancer cells.Typically, prostate cancer cells develop resistance to abiraterone. Consequently, disease progression and treatment failure is observed in roughly a year (11 months for Prostate Specific Antigen [a biomarker] increase and 16 months for radiographic progression). While small, the study shows that the median time to PSA and radiographic progression is at least 27 months when patients follow the model derived treatment schedule.

Adaptive therapy is designed to take advantage of the "cost of resistance." When abiraterone concentrations are high, it is hypothesized that the relative cost of resistance is lower than the fitness benefit, so cancerous cells are likely to evolve into abiraterone resistant strains. Adaptive therapy aims to use treatment holidays to protect against the development of these resistant strains. In general, treatment is suspended when PSA reaches half of pre-treatment levels, and only restarts once pre-treatment PSA levels have been reached.

I'm going to conflate mathematics with statistics as Carlo Beenakker did. Then the earliest application that I know of is that Decision Trees were invented by Breiman et al. to analyze the issue of- among people who have had a heart attack, which patients were most likely to have another heart attack. I also believe that Ed Frenkel, in his autobiography, claimed to have developed a similar methodology to explain to doctors how to triage. A really beautiful application is something called James-Stein estimation which deals with the inadmissability of the mle for estimating (true) means of many variable. The short story is that if you have 4 or more series of observations, then the individual estimated mean of each series of observations is not the best estimates of the real means. This was applied by Efron (mentioned in various books and papers including "Large Scale Inference") to the question of deciding which genes (via gene expression) were likely to be influential in causing a specific cancer.