Beyond clinical guidelines: mathematical models of treatment protocols for decision support and in silico clinical trials
Paeon Virtual Doctors are mathematical models precisely defining the decision strategies followed by physicians when administering given treatment protocols.
On the basis of the protocol to be followed and actual patient clinical data, a Paeon Virtual Doctor defines the next actions to be performed, as timing and amount of drugs to be administered as well as the day of the next visit.
Paeon Virtual Doctors go much beyond usual clinical guidelines. They precisely encode the strategy to follow on the basis on patient data, as external factors (for example, age) and clinical measurements (for example, as for fertility treatments, antral follicle count and hormone blood levels).
The methodology developed within Paeon allows medical scientists to define Virtual Doctors encoding treatment protocols in all areas of medicine.
During the project lifetime, we have successfully developed Virtual Doctors encoding the three main fertility treatment protocols currently in use at the Reproductive Endocrinology division of University Hospital Zurich.
A wide family of treatment protocols can be formalised by the common top-level decision structure shown in the figure.
Typically, treatment protocols start with a planning phase, where initial decisions are taken on the basis of patient data, which can be external factors and/or clinical measurements.
After the Planning phase, treatment protocols typically proceed with a loop structure that, at each iteration, requires to take some clinical measures on the patient (for example, via blood samples) and to decide, on such a basis:
At each iteration, a treatment protocol may also require to check whether some safety conditions are satisfied.
Safety conditions must always hold to guarantee patient health.
For example, to reduce risk of treatment adverse effects, it is important that some patient measurements do not exceed given thresholds. In case of violations, the treatment must stop immediately with a failure.
At each step of the process, a treatment protocol also checks whether the treatment success conditions (or goals) have been met and, if not, whether there is still hope to achieve such conditions.
If there is evidence that treatment goals cannot be attained any more, the treatment protocol ends with a failure.
Finally, at each step of the process, the treatment protocol updates and stores its state and waits until the prescribed day of the next visit.
A formalisation of a treatment protocol as an instance of this general top-level structure is called a Virtual Doctor.
In fact, a Virtual Doctor precisely defines the strategy that an actual physician would follow when administering that treatment protocol.
In particular, a Virtual Doctor, at each step, performs all the activities in the figure, which lead in the automated decision of the next clinical actions to be performed on the patient (namely: prescriptions of drug dosage and additional measurements, and decision about the day of the next visit).
Paeon considered three fertility treatment protocols currently in use at University Hospital Zurich (UZH, Paeon partner): Long, Short and Antagonist protocols.
The three protocols consist of different phases. They differ in applicability criteria, overall duration, structure and drugs used.
The first phase of the Long protocol is the Planning phase.
Its goal is to decide the duration of the subsequent Downregulation phase and the initial drug dosage for the forthcoming Stimulation phase.
Such decisions depend on several factors: patient age, causes of infertility, diseases (for example, PCOS) and other external factors (like obesity).
Also, transvaginal ultrasounds are performed to compute the patient antral follicle count and blood samples are taken to measure the level of different hormones.
The Downregulation phase of the Long protocol has a duration varying from 3 to 4 weeks which is decided during the previous planning phase.
During Downregulation, the patient is treated daily with GnRH analogues (like Triptorelin).
The goal is to suppress the natural menstrual cycle.
At the end of Downregulation, if Estradiol blood level is below a certain threshold, the Stimulation phase can start.
The goal is to stimulate a large number of follicles to grow and reach maturity.
To this end, the patient is treated daily with a drug based on a mix of FSH, LH or hMG plus GnRH analogues for 2-3 weeks.
The initial drug dosage is decided during the Planning phase. However, dosage is adjusted every a few days depending on how the patient is reacting. This is evaluated by deciding to take additional clinical measurements (hormone levels via blood samples and sizes of the growing follicles by means of transvaginal ultrasounds).
During this phase, a continuous monitoring of several safety conditions is performed as well, in order to keep risks under control.
The Stimulation phase terminates when the treatment success condition on the number and size of the dominating follicles is achieved, or when there is no hope to achieve success within a given deadline.
In case of successful treatment, the Ovulation Induction phase is performed, which consists of a single administration of a drug based on r-hCG aimed at inducing ovulation.
After ovulation induction, occytes are collected and fertilised.
The Short and Antagonist protocols differ from the Long protocol in that:
We have validated our formalisations of the three Virtual Doctors defining the Long, Short and Antagonist protocols currently in use at UZH against retrospective clinical data.
The goal is to evaluate how precisely the decisions of our Virtual Doctors match those of the actual physicians.
The figure shows the differences between the drug dose suggested by the Virtual Doctor formalising the Long Protocol during the 5 decision points of the Stimulation phase and the corresponding doses reported in retrospective data as being actually administered by the physician. Dose differences are expressed in multiples of dose quanta (37.5 IU). As typical doses for stimulation drugs range from 150 to 400 IU, a difference of 1-2 dose quanta has to be considered very small.
The result is that the Paeon Virtual Doctor for the UZH Long protocol takes decisions which differ from those actually taken by physicians for at most 1 (resp. 2) dose quanta in 66% (resp. 92%) of the cases. Similar results apply to the Virtual Doctors for the Short and Anta protocols.
Overall, this shows that Paeon Virtual Doctors accurately formalise the decision strategies of actual UZH physicians. Results can be further improved by including, in the formalisations, additional aspects, as physician knowledge about the outcome of the patient’s previous treatments.