Dynamical digital biomarkers

Smartphones and wearables continually capture objective and subjective data that can be turned into clinically meaningful information: into a digital biomarker.

What would these digital biomarkers be used for? The data is collected by the smartphone. It can be processed on the smartphone. And hence the result should also be directly displayed on the smartphone.

Think about the following: Acute medication is often started too late. Prevention needs to be over a sufficiently long time with often daily use of drugs with adverse side effects. What if there would be a third mode that takes the best of both worlds and avoids their worst characteristics: a preemptive therapy? Your smartphone tells you early enough when and which drug to take, but not so often that adverse side effects become a problem.

Taxonomy of biomarkers

Let us first start with a simple taxonomy of biomarkers. This can be provided by firstly dividing them into whether they are obtained before or after the onset of the disease. In other words, therapeutic interventions based on biomarkers can be preventative or acute, with an intermediate »preemptive« phase.

Before the onset of a disease, susceptibility and risk biomarkers determine preventative interventions. After onset of the disease, a biomarker can be diagnostic or monitoring. Positioned between these phases are the dynamical biomarkers—at least one type of them, two more types can be present after the onset. (See notes on terminology.)

In the remainder, we focus our full attention on the dynamical biomarkers. However, it should also be mentioned that all aforementioned biomarkers in these phases can be prognostic, predictive or are true companion diagnostics. So what kind of prediction is made? A positive prognostic biomarker reveals that the condition is such that an intervention can help. A positive predictive biomarker reveals that a particular intervention under consideration can help. And finally, companion diagnostic predict the intervention’s efficacy or its »toxicity«. These are aspects of a taxonomy that should be taken into account even if they are only marginal referenced to in the following.

Early-warning signs come in three types

The subgroup of dynamical biomarkers serve as early-warning signs. These signs indicate an imminent critical transition in an illness and therefore determine the optimal point in time for medical interventions.

Three applications of dynamical biomarkers must be distinguished: early disease detection (e.g. cancer), early disease progression (e.g. stages of multiple sclerosis or Parkinson’s disease) and the early detection of episodes in chronically ill patients (migraine attacks, seizures in epilepsy, transient ischemic attack in cardiovascular disease patients). In a single disease, all three types of biomarkers can exist.

Turning patient-generated data in actionable information for treatment

This article deals with dynamical digital biomarkers (DDB), that is, I consider how we can turn patient-generated data in actionable information for evidence-based treatment decisions. In the digital domain, only the latter two types, i.e., detecting progressive stages (DDB type II) or intermittent episodes (DDB type III) of patients already being diagnosed with a chronic disease seem to be of greater medical interest. So we will focus on these to types.

The prime example is migraine

Take as the prime example migraine. Migraine is a chronic disease. We can distinguish two stages, namely the episodic and chronic form. Note that one must not confuse the stage of »chronic migraine« with migraine being a chronic disease. »Chronic migraine« refers to the chronification of pain, so that the physiological origin of the pain changes from nociceptive pain to central pain, which results in different targeted therapeutics. At least in theory—the exact nature of chronic migraine remains a scientific debate (see below). There is no doubt, however, that a migraine patient has a chronic disease independent which of the two forms of migraine he or she suffers from.

Every migraine patient tries to find putative triggers

Simply speaking developing a DDB type III of migraine is nothing but enabling a migraine patient to predict his or her attacks. Every migraine patient tries this.

Almost every migraine patients tends to carefully investigate signs of the imminent attack, for example changes in the environment (weather or light conditions) and lifestyle events (skipping a meal, doing sports) and changes in vital signs, indirectly assessed (e.g. by yawning or abdominal discomfort) or directly measured with the help of a wearable.

Usually this quest is formulated as the search for triggers. But studies on migraine suggest that early warning signs are easy to confuse for trigger factors.

“Although migraine sufferers often are convinced that certain food, stress, bright light, neck pain, and other factors may trigger attacks, under controlled experimental conditions, there is very little if any evidence that these putative trigger factors can actually provoke attacks. Instead of being the trigger initiating an attack, craving certain food, perceiving normal events as stressful or normal light intensities as too bright, and experiencing neck pain in the few hours to days prior to the clinical manifestation of a migraine attack more likely are early premonitory symptoms of an attack.” (Dahlem et al. Cephalalgia 2015)

So if a migraine patient asks: what triggers my migraine?, and uses a migraine app to get an answer, what he or she actually asks for is a DDB of type III.

It is noteworthy that confusing body signs for triggers is reminiscent of confusing stress (internal physiological state) for stressors (external stimulus). However, switching in terms or between the internal and external perspective should be avoided when discussing DDBs of type III, because this simple realization does not help.

The actually startling situation is another one: Feeling a little bit of stress is actually helpful. But at a certain point there is a reduction of benefit due to stress-related state changes called allostatic load. The allostatic load has individual characteristics. A DDB of type III should provide an estimate of the acute allostatic challenge derived from patient-generated data. Thus a DDB of type III in migraine can and must be understood »through the lens of maladaptive stress responses«.

The situation is in fact a bit more twisted. If you have a biomarker, you need to know sensitivity and specificity. But this is not that simple in DDB of type III, because such a sign is a warning sign to be act upon.

Of course in prognostic clinical trials one could try not to reveal the information. But since the DDB of type III is generated by patient-driven data one might question how well this really works. A subset of the data streams going into the DDB of type III are premonitory symptoms that the patient is aware of, so that this cannot be fully blinded.

In the following illustration, the situation for one month is displayed with a biomarker that takes the presence of a postmonitory symptome into account, like yawning or neck pain. Does excessive yawning predict a migraine attack in a patient? And if so, does it predict a migraine attack more reliably than back pain?

30 vertical bars indicate the days of one month. Headache attacks occur on four days, marked yellow if correctly predicted (true positive) and purple if not (false negative). Other combinations are more interesting, in particular a false positive day (marked blue) might be actually a true positive day one which the attack was successfully suppressed due to the information of the biomarker.

In fact the whole consideration repeats as one has to consider always a particular premonitory symptom (PPS), like excessive yawning, and other signals that compromise a particular DBB type III out of a whole set of data sources.

In other words excessive yawning can be prognostic: answering whether an attack might be successfully suppressed. Or predictive: answering whether an attack might be successfully suppressed only by a certain intervention. Think for example that the presence of neck pain might indicate that massage will help but drinking coffee does not.

Ten questions about dynamical digital biomarkers type III

So beside the information on sensitivity and specificity there are eight more questions that should be asked about dynamical digital biomarkers type III:

(1) What is the probability of being alarmed? (sensitiv)
(2) What is the probability of having no false alarm? (specific)
(3) What is the probability an alarm is not false? (precise I)
(4) What is the probability no alarm means no attack? (precise II)
(5) What is the probability the alarm/no-alarm is correct? (accurate)

And five more if there is a correctly or incorrectly presumed premonitory phase …

(6) … what is the probability of a DDB type III being in the actual (correct) phases? (sensitiv)
(7)… what is the probability of a DDB type III being in the incorrect phases not followed by headaches? (specific)
(8) … what is the probability of a DDB type III being correct? (precise I)
(9)… what is the probability of the absence of the DDB type III being correct (precise II)
(10)… what is the probability of presence and absence being correct (accurate)

The quest has just begun

With the advent of sophisticated data collection with migraine apps the quest for dynamical digital biomarkers has only begun. But once it is clinically proven, a DDB of type III opens up preemptive therapy strategies that unite the best of both worlds of acute and prophylactic therapies.

For example, recent advances in migraine therapy will bring so-called »ditans« on the market. Ditans are acute treatments options similar to triptans but seemingly without having vascular effects and a lesser risk of medication overuse headache. So unlike triptans, ditans can be taken in preemptive mode, if we have a dynamic biomarker predicting the imminent attack.

Migraine is the prime example of chronic diseases with episodic manifestations (CDEM). And a DDB of type III is only one example of how a well thought through digital transformation strategy in chronic disease care can change the way we digitally treat these diseases in the future.

What about disease progression?

What remains is a short view on DDB of type II, that is, detecting early disease progression in terms of stages. For chronic diseases with episodic manifestations, there is a surprisingly simple answer: if there exists a proven DDB of type III in the low frequency episodic phase and the disease is progressive, that is, the attack frequency increases, it is to be expected that the DDB of type III signal breaks down.

This breakdown itself, maybe in combination with other new symptoms, could be a DDB of type II, that is, an early prediction of the new disease stage.

In the case of migraine, this would mean that once attacks become less predictable and new symptoms arise, such as for example allodynia, this could be a dynamical biomarker for chronic migraine. Such a definition for chronic migraine at least seems to be more useful for early targeted therapeutics than the current rather artificial threshold of remaining headache free for at least less 50 % of the days in three months.

In other words, the digital transformation strategy in chronic disease will not only improve diagnostics but actually change it from a symptom-based one to an etiological catalyzing the success of precision medicine.

Notes

Why »dynamical biomarker« and not »dynamic biomarker«? While both spellings exist, the former is a hint at the related »dynamical systems theory«. This spelling was allegedly introduced at the »Heart Rate Variability 2006:
Techniques, Applications, and Future Directions« . Furthermore, dynamical biomarkers are also related to the concept of »dynamical network biomarkers« (DNB), which where introduced in migraine in 2013.

• M. A. Dahlem, S. Rode, A. May, N. Fujiwara, Y. Hirata, K. Aihara, J. Kurths, Towards dynamical network biomarkers in neuromodulation of episodic migraine, Translational Neuroscience, 4,282–294, (2013)
• M.A. Dahlem, J. Kurths, M. D. Ferrari, K. Aihara, M. Scheffer, and A. May, Understanding Migraine using Dynamical Network Biomarkers, Cephalalgia, (2014).