Abū ʿAlī al-Ḥusayn bin ʿAbdullāh ibn al-Ḥasan bin ʿAlī bin Sīnā al-Balkhi al-Bukhari is the formal name of Ali ibn Sina, known in Europe with his Latinised name, Avicenna. In 1025, he completed The Canon of Medicine, an encyclopedia of medicine in five books, which remained a fundamental reference in any decent medical library for centuries, in one of the many translations from Farsi (Persian language).
However, one element of the Canon survives nearly untouched after 1000 years: how to assess the efficacy of a new treatment through experimentation, in short, clinical trials. The seven conditions Ibn Sina proposes are purity of the drug, testing for only one disease, use of a control group, use of dose escalation, need for long-term observation, need for reproducible results, and the need for a human in addition to animal testing. These are pretty much still the cornerstone concepts of contemporary clinical trial design.
The epistemological perspective of Ibn Sina was what we would today refer to as frequentist inference. This worldview, mainly rooted in social sciences, relies on a few fundamental assumptions (epistemologists will forgive me a certain degree of oversimplification):
- Experimental observations can rarely be quantitative and are always biased (i.e., affected by systematic errors).
- Truth exists but is hidden.
- Experimental observations are collectively informative, whereas prior knowledge is considered scarcely informative; decisions are based solely on the likelihood of observing the data under various assumptions.
- Decisions are reduced to a binary choice between H0 (null hypothesis), which is the choice we would make in the absence of any experimental observation, and the alternative hypothesis H1, which is the opposite of H0.
- By default, we assume H0 to be true; we change our decision to H1 if there is overwhelming experimental evidence that H0 is false.
The key concept here is prior knowledge. Prior knowledge is what we believe to be true before we start the experimental observations. There are other domains of knowledge, such as physical sciences, where it is believed that we can provide causal explanations in the form of prior knowledge for some manifestations of reality. In such cases, prior knowledge can indeed be informative. This worldview can be summarised in the Bayesian inference, which assumes:
- Most experimental observations are expected to be quantitative, and our measurement methods are nearly free of systematic errors.
- Truth does not exist: there are beliefs (hypotheses that someone believes are true) that are more or less probable.
- Every hypothesis has a posteriori probability, defined as the product of the probability derived from the a priori knowledge available (prior) and that derived from experimental observations (likelihood).
- The hypothesis with the highest probability is used to make a decision.
The fact that two fundamentally different and largely incompatible worldviews coexist in science has always fascinated me. But it makes sense: in social science, very little can be "measured", and no one expects any "law of sociology" to exist any time soon (Interestingly, this has not always been the case: look at the concept of
psychohistory in Isaac Asimov's science fiction novels). On the other hand, in physics and engineering, we trust we can observe experimentally natural phenomena in ways that yield reproducible, reasonably unbiased quantifications. From these, we can formulate sufficiently universal causal explanations that, when resisted by extensive attempts of falsification, are considered "laws of physics". Within their limits of validity, we expect the laws of physics to be universally true; thus, we can use them as reliable prior knowledge in decision-making.
The birth of modern physics is usually placed between 1543 (publication of Copernicus' "De Revolutionibus Orbium Coelestium" on heliocentric model) and 1687 (publication of Newton's "Philosophiæ Naturalis Principia Mathematica", unifying law of motion and universal gravitation). Maybe being biased by my being Italian, I like 1638 (publication in Leiden of "Discorsi e dimostrazioni matematiche intorno a due nuove science", The Discourses and Mathematical Demonstrations Relating to Two New Sciences, Galileo Galilei's final book. Whatever the date, this is some six hundred years after ibn Sina canon. And, to be fair, Galileo's new science dealt mostly with falling stones, which are much simpler than living organisms.
Thus, I dare to say that ibn Sina was right in conceiving the testing of new treatment as the purest frequentist investigation. So were the many medical researchers followed him through the centuries, always using the same approach. For centuries, investigating human health and diseases was quite close to social science: little could be measured, every observation was inevitably biased, and prior knowledge had little informative value.
I teach my students that the music changed in 1943 when renowned
physicist Erwin Schrödinger was invited to give a course of public lectures at Trinity College in Dublin. Schrödinger was already world-famous for his quantum theory work, so it surprised everyone when he announced that the course would be titled "What is life?". In 1944, he published a book from these lectures entitled "
What Is Life? The Physical Aspect of the Living Cell". Schrödinger's ideas apparently had zero impact on biomedical researchers. Still, the seed was planted: "How can the events in space and time which take place within the spatial boundary of a living organism be accounted for by physics and chemistry?". Or, in our words, can we use the laws of physics and chemistry as prior knowledge to investigate the efficacy of new treatments better?
It was a long and winding journey from
Schrödinger's book to the
Avicenna Research Roadmap, which postulated the possibility of using
in silico methodologies to better investigate the efficacy of new treatments by leveraging on a large body of prior knowledge developed by physiologists, biophysicists and bioengineers, starting from the seminal work of
Hodgkin and Huxley on the action potential of neurons, published in 1952. But today, the body of prior knowledge about the causation of human pathophysiology phenomena which resisted extensive falsification is substantial. Why should we not use it also for the assessment of new treatments?
The clever combination of computer modelling and clinical experimentation offers alternatives to animal experimentation, the reduction of the cohorts to be enrolled in clinical trials, a drastic reduction of the attrition rate (number of new drugs that fail to show safety and efficacy), reduce drastically the time to market, and the costs of development. Why are we resisting such innovation?
My explanation is that most experts working for drug regulatory agencies share the classic frequentist inference worldview; moving to a Bayesian inference worldview is a vast cultural change, a true cultural revolution that scares many practitioners. But as we celebrate the 1000 years of the Canon of Medicine, I beg all stakeholders to stop postponing this radical change. Like all revolutions, it will be painful and ridden by mistakes, but ultimately, it is worth the risk.
