| Why Personalized Medicine?
Personalized medicine is a concept and a practice as old as medicine. Advances in clinical biology, imaging and knowledge management (Internet, AI, M&S) have made it more feasible and possibly more effective. This blog focuses on how technology such as AI and mechanistic modeling can support tailored treatment decisions for patients. It does not cover screening which comes with challenges of its own. Definition: Personalized medicine (PM) involves selecting the most effective therapeutic strategy for a patient, tailored to his/her specific characteristics and needs and agreed upon with him/her. Snapshot of the medical decision process:
Apply decision theory – The process can be summarized in three steps:
Therapeutic objectives and prediction of benefit (to what extent the therapeutic objectives can be achieved) are the two main drivers of decision theory in medical practice. Role of Technology:
Physician’s approach:
|
History hallmarkSince Hippocrates of Kos (circa 460 BC), the care provided by a caregiver must be adapted to what his patient is suffering from. The first mention of the term “personalized medicine” is difficult to identify. In more recent days, personalized medicine was defined as “the prescription of specific therapies that are best suited to an individual based on pharmacogenetic and pharmacogenomic information.” At the time, it was essentially a question of adjusting the doses of drugs whose metabolism was dependent on genes whose variants were correlated with abnormal plasma concentrations, accompanied by adverse effects or inefficacy. At the end of the previous century, Canadian academics invented evidence-based medicine, which consists of using current scientific data to meet the specific needs of each patient. Certainly, they advise the use of “clinical common sense” to adjust – transpose – current scientific data to these needs. This approach has come up against the considerable amount of current scientific data and the challenge of evaluating its validity. Then, since the beginning of the new century, the exponential growth of genetic epidemiology studies, based on new techniques for exploring the genome – sequencing – and the beginning of proteomics, instilled a prognostic dimension. These were, for example, the discovery of the association of the polymorphism of the estrogen receptor gene ER-alpha and the prognosis of breast cancer. The definition of personalized medicine later expanded but retained its genetic legacy. Is it the publication in 2002 of the observation that an algorithm integrating the expression of 70 genes predicts the prognosis of women with breast cancer better than a traditional clinical score? Is it the influence of pharmacogenetics…? The fact remains that personalized medicine is associated with genes and their expression. Even today, for many, personalized medicine is only possible based on genome exploration. In 2013, the Australian National Health and Research Council published a note: “personalized medicine (also known as stratified medicine or precision medicine) relies on genetic knowledge” (i.e. the links between genes and diseases and the effectiveness of certain therapies) to predict the occurrence of disease. to help with lifestyle choices or to adjust treatment to an individual”. Then came AI. From the beginning of the Covid-19 epidemic, the AI community, in particular, has been quick to develop software to allow hospitals to diagnose or triage patients faster. Finally, several hundred predictive tools were developed. None of them made a real difference to the traditional approach for Covid suspected patients, and some were potentially harmful. But these failures can be explained by the tiny amount of data available to calibrate the algorithm. The author concluded: “the pandemic could help make medical AI better”. In the field of medical imaging, where data is available in large quantity (in the form of pixels), the usefulness of AI is no longer in doubt. |
Warning: The thoughts on personalized medicine below do not concern screening, which poses specific problems. We place ourselves here in the situation of a physician who must face a patient’s complaint or disturbing symptom.
Hence the definition: personalized medicine (also coined as precision medicine) consists of the physician choosing for a subject the therapeutic or preventive strategy most likely to achieve the therapeutic objective set in concert with him.
Several wordings are used to name personalised medicine; some, such as 4P medicine, deepening the meaning more than it used to be.
Let’s start with a theoretical step flow of the doctor’s approach viewed through the prism of decision theory. The doctor begins by identifying the patient’s complaint, then examines it to record all the symptoms, from which his knowledge of the diseases will lead him to a diagnosis. Based on the diagnosis, the doctor looks for the prognosis, i.e. the probabilities that the patient’s disease will evolve towards this or that more or less severe outcome. Then, taking the patient’s preferences, the doctor prioritizes the outcomes that should be avoided. Once these therapeutic objectives have been set, the doctor makes a list of the therapeutic means available for preventing each outcome. For each therapeutic objective/therapy pair, he predicts the intensity of the benefit – i.e. the absolute benefit, not the relative benefit – that his patient will derive from taking the therapy and the disadvantages that he may have to face (cost, burden, side effects). The absolute benefit is the amount of gain on the probability (e.g. of death) or the quantity (e.g. of pain) of outcome that is expected from the therapy. Applying the notion of threshold to each predicted absolute benefit, the physician ranks possible therapies by decreasing amount of predicted benefit above the threshold, specific to each pair therapeutic objective/therapy. Eventually, based on this list and exchange with the patient, the doctor prescribes.
The quantitative prediction of the benefit that each therapy could bring for a therapeutic objective, adjusted to the characteristics of the patient, is shown to be possible by the effect model theory.
In some cases, well illustrated by the situation of hypertension or hypercholesterolemia, for which no amenable pathological consequences are yet detectable for the patient at the time of the consultation (after appropriate additional investigations), the physician will move directly from the symptom to the therapeutic objective. The diagnostic step is replaced by the observation of the absence in this patient of pathology associated with what is called a risk factor (in this case high blood pressure or high blood cholesterol).
Another situation deserves to be mentioned. The therapeutic objective is far away, but there is a physiological, biological or imaging criterion that is closer in time, which will serve as a temporary substitute for the clinical therapeutic objective. The physician postulates, and warns his patient, that if this intermediate objective is not achieved, it will be necessary to amend the therapeutic decision, perhaps simply by changing the therapy regimen.
The physician’s approach was presented above as linear. In practice, this is rarely the case, with back and forth, particularly at the diagnostic stage.
Decision theory is particularly applicable to the physician’s approach and allows it to be summarized in three stages: 1) Establish the diagnosis and prognosis; 2) from which are the possible therapeutic objectives and therefore the possible therapies are listed; 3) choice of a therapy based on predicted benefits.
The disease has current and future manifestations, so this dynamic is the guiding principle of the physician’s approach as described above. It is fueled by knowledge about this disease and available treatments for it, knowledge that is either in the doctor’s head or easily accessible by one material means or another.
Described in this way, this approach is perfectly personalized since its course depends on the characteristics of the patient – clinical, biological, anatomical symptoms (obtained by imaging), age, sex, weight, etc.
So why are we talking about personalized medicine today?
Reasons for this are to be found in a rather old observation – physicians cannot access and use all relevant available knowledge – and in two recent technological revolutions that are intertwined: that of the means of exploration in clinical biology – with the different omics – and imaging, and that of knowledge management and exploitation. This latest revolution is based on AI and mechanistic modeling and simulation (M&S or AI-QSP).
Divers AI modalities (such as deep learning algorithms, convolutional neural networks, generative adversarial networks,…) have been tested to help with diagnosis. Some are routinely used in medical imaging. AI speeds up the interpretation of images (there are 4,000,000 pixels in a standard resolution chest X-ray) and allows early detection of diseases in radiology, pathology, cardiology. AI-based image processing facilitates personalized treatment decision.
Beyond facilitating image based diagnosis, AI presumably lends itself to diagnostic assistance because, starting from the patient’s symptomatology, it makes it possible to search knowledge bases and databases for identical or very similar profiles of patients for whom the diagnosis has been established.
On the other hand, for the therapeutic decision, AI is not the best option. Of course, the AI would be able to find identical profiles of patients who have received a given therapy and provide insights on the evolution of the doctor’s patient, as well as the comparison with the evolution with other therapies for patients with very similar profiles. But this is on condition that the therapy is not brand new. Further, these comparisons of predictions of benefits with the available therapies suffer a severe limitation: they are not randomised, making the management of bias rather uncertain. Yet, the main limitation of the AI in the therapeutic decision is the impossibility of taking into account hidden characteristics that the doctor cannot inform and which could be impactful on therapy efficacy or toxicity. This leads to uncertainty about the adequacy of the profiles found to meet that of the patient. This adds to the problem of the quality of the data available in the accessible databases. And, related to the data quality issue, the problem of the age of the data and the relevance of the diagnoses and outcome definitions.
Once the diagnosis has been made, the mechanistic models of the disease and of the available therapies will make it possible to predict the benefits for the patient, taking into account his hidden characteristics, the evolution without and under considered treatments, allowing to compare the benefits for each therapeutic objective/therapeutic pair. This is done either by rotating the models for each patient, or by using the effect models of each therapy for each therapeutic objective identified beforehand.
The AI-QSP approach is not conducive to diagnostic assistance because of the number of models that would have to be available in the library. The diagnosis is on the other hand a static operation while the contribution of the AI-QSP is to represent the dynamics of the modeled processes.
Perhaps more than for other stages of medical decision-making, the prediction of prognosis by AI or M&S comes up against the obstacle of hidden patient characteristics.
The AI approach does not offer any specific solution, M&S approach integrates such a solution. The AI techniques available today do not allow these hidden characteristics to be extracted from the data (they are not recorded in the databases) and knowledge (they often emerge at the knowledge management step) on which they are based. This leads to uncertainty, or even to the possibility of error regarding the list of therapeutic objectives and their probabilities. This uncertainty and errors are not manageable by a specific algorithm. Whether traditional statistical techniques could be of help remains unfixed.
The disease dynamics carried by the disease model in the M&S approach makes it possible to predict a distribution of the probability of a therapeutic objective from the plausible distribution of hidden characteristics represented by the virtual population. Uncertainty exists but it is manageable, thanks to a specific algorithm.
To carry out this stage of prediction of the prognosis, we propose to use the two approaches and to confront the results obtained: the list of therapeutic objectives and their probabilities.
Apart from the technical problems mentioned above and those of acceptability by caregivers and patients, the personalization of medicine comes up against the rigid framework of good clinical practice guidelines and in some countries such as France the rules for the reimbursement of health expenses.
