Causation in Systems Medicine: Epistemological and Metaphysical Challenges Jon Williamson 8 January 2015 EBM+ workshop, Kent
Contents 1 Systems Medicine 2 Epistemological Challenges for Systems Medicine 3 Metaphysical Challenges for Systems Medicine 4 Epistemic Probability for Evidence Integration 5 Epistemic Causality to account for SM Methodology Bibliography
1 Systems Medicine Systems biology studies biological systems holistically. Typically, systems of molecules and their causal interactions within the cell. Using data-intensive functional genomics techniques. EG Transcriptomics, metabolomics, proteomics. Appeals heavily to mechanistic knowledge. Used for mechanism discovery. Systems medicine applies systems biology to medicine, with two goals: Practical. For diagnosis, prognosis and treatment. IE Causal discovery. Theoretical. For discovering pathophysiological mechanisms. NB Includes higher-level clinical and environmental data and mechanisms.
Mechanistic models vs causal models Causal models look like this: X W Z Y Normally, if X causes Z then: X and Z are probabilistically dependent conditional on Z. (Except in cases of overdeterminaton...) There is some mechanism by which X leads to Z. (Except with absences, double prevention...)
Mechanistic models look like this:
The promise of systems medicine: Big data: More robust conclusions. Increased personalisation. Can discover effects that are washed out in a large population. Treatments can be better targeted. Can handle more complex pathophysiological mechanisms. Opens up a richer seam of causal claims. NB Lots of promises bandied about. EG SM is P4 Medicine : predictive, preventative, personalised and participatory.
We stand at the brink of a fundamental change in how medicine will be practiced. Over the next 5-20 years medicine will move from being largely reactive to being predictive, personalized, preventive and participatory (P4). Technology and new scientific strategies have always been the drivers of revolutions and this is certainly the case for P4 medicine, where a systems approach to disease, new and emerging technologies and powerful computational tools will open new windows for the investigation of disease. Systems approaches are driving the emergence of fascinating new technologies that will permit billions of measurements on each individual patient.... We predict that emerging technologies, together with the systems approaches to diagnosis, therapy and prevention will lead to a down turn in the escalating costs of healthcare. In time we will be able to export P4 medicine to the developing world and it will become the foundation of global medicine. The democratization of healthcare will come from P4 medicine.... It is evident that the business plans of every sector of the healthcare industry will need to be entirely transformed over the next 10 years (Galas and Hood, 2009, p. 1)
Challenges: More observational evidence. Confounding. Depends heavily on mechanistic background knowledge. Danger of error. Big data. Computational complexity. Many simplifying assumptions. Data / evidence integration. Metaphysical. Hard to find an account of causality that fits SM practice. SM: Demands a lot in understanding / so many challenges.
2 Epistemological Challenges for Systems Medicine the technologies to get large amounts and different types of data will soon be affordable and readily available in the clinic. But what are we going to do with these long lists of data? Taking all this data into account, and integrating it, is not a trivial task when taking decisions in the daily practice. The sheer volume of data necessitates multidisciplinary interaction; a general practitioner cannot make diagnostic and therapeutic decisions based on hundreds of thousands of data points of -omics data by integrating it in his or her head, they require support of experts from other fields. The development of mathematical and information science tools has opened up possibilities to mine these large sets of data, to postprocess them and to reduce the noise in the data... There is a need for flexible, integrative systems approaches to combine such -omics data with clinical, societal and environmental factors including sex, type of work, sleep and eat habits, etc. (Vandamme et al., 2013, pp. 892 893.)
There is a need to find a model that fits multiple datasets. Each kind of data yields a fingerprint. IE A model that gives a partial indication as to what is causing what. EG Metabolomic data. EG Proteomic data. EG Transcriptomic data. EG Clinical data. EG Patient-reported outcomes. A model is needed that fits all these sources. The handprint. There is no consensus as to how to do this.
Worse than that: This overall model also needs to fit mechanistic evidence. NB There is a tendency to think that if it s not a dataset then it s not evidence. But it s clear that mechanistic evidence guides SM at all stages. Intuition / eyeballing the mechanistic evidence isn t a realistic prospect. There is too much of it in the SM process. Need to make its contribution explicit. Need to be explicit about the role it plays: EG To rule out the hypothesis that a correlation is non-causal. EG To help determine the direction of causation, or whether a correlation is due to a common cause. Need to make quantitative prediction across levels of mechanisms.
3 Metaphysical Challenges for Systems Medicine Does any metaphysical theory of causality fit with the epistemology of SM? If causation is just dependence, why does SM use evidence of mechanisms to justify causal claims in cases where the data that provides the full structure of associations? If causation is just mechanistic connection, why does SM use evidence of associations to justify causal claims in cases where the mechanisms are known?
4 Epistemic Probability for Evidence Integration One way to meet the epistemological challenges might be to apply objective Bayesianism: ω 1 E P P = ω 2 P ω 3
Why? Probability. Degrees of belief should be probabilities. Avoid sure loss. Calibration. Degrees of belief should fit evidence. EG Her degrees of belief should be set to physical probabilities where known. Avoid almost sure long-run loss. Equivocation. Degrees of belief should equivocate between basic possibilities. Minimise worst-case expected loss. (Williamson, J. (2010). In defence of objective Bayesianism. Oxford University Press, Oxford.) An integrated justification: Controlling worst-case expected loss requires that degrees of belief should be probabilities, calibrated to physical probabilities, and sufficiently equivocal. (Landes, J. and Williamson, J. (2013). principle. Entropy, 15(9):3528 3591.) Objective Bayesianism and the maximum entropy
From fingerprint to handprint: P E 1 P 1 E 2 P2
Objective Bayesian nets A Bayesian net that represents the objective Bayesian probability function. IE The maximum entropy probability function. This is the handprint model.
Dataset 1: Dataset 2: Bayesian net 1: A B C Jill Keith Linda............ Bayesian net 2: C D E Jim Kirsty Lionel............ A B C D C E Objective Bayesian net: D A B C E
Mechanistic structure: recursive Bayesian nets M = DNA damage response mechanism S = survival after 5 years M S P(M), P(S M) M = m 1 : M = m 0 : D R D R P m1 (D), P m1 (R D) P m0 (D), P m0 (R)
Inconsistent fingerprints Convex hull approach: E 1 P 1 E 2 P2 P
Higher-order approach: P E 1 P 1 E 2 P2
Convex hull approach: E 1 P 1 E 2 P P 2 E 3 P 3
E 2 P 2 Higher order approach: E 1 P 1 P E 3 P 3
5 Epistemic Causality to account for SM Methodology Causality is a feature of the way we represent reality rather than out there (Williamson, 2005, Chapter 9): Irreducible to patterns of difference making / mechanisms. Our causal beliefs help us to predict, to explain and to control reality. We have these causal beliefs because of this inferential utility. Not because there is some non-epistemic causal relation that is the object of those beliefs. Analogy with the trihoral relation: C C G P O L
Analogy with epistemic probability: Both are a kind of belief, rather than a belief about: Epistemic probability: a probabilistic belief is a degree of belief of the form P(E) = rather than a belief about a worldly probability. Epistemic causality: a causal belief is a directed belief of the form C E rather than a belief about a worldly causal relationship. Both are explained in terms of inferential success: We have probabilistic beliefs in order to help us bet on events. We have causal beliefs to draw PEC-inferences. Both can account for apparent objectivity: The probabilities are the degrees of belief you ought to adopt if your evidence were to include all fundamental matters of fact. The causal relationships are the causal beliefs you ought to adopt if your evidence were to include all fundamental matters of fact.
Norms Objective Bayesianism: Probability. Your degrees of belief should be probabilities. Calibration. Degrees of belief should also be calibrated to evidence of frequencies. Equivocation. Degrees of belief should otherwise be no more extreme than the evidence demands. Epistemic causality: Acyclicity. Your causal beliefs should be acyclic. Calibration. Causal beliefs should also be calibrated to evidence. Dependence. If E is not overdetermined and C E then there should be evidence of a chain of dependence from C to E. Mechanism. If C and E are not absences and C E then there should be evidence of a chain of mechanisms from C to E (except in cases of double prevention).... Equivocation. Causal beliefs should otherwise be no more extreme than forced by evidence.
Conclusion Systems medicine promises a lot but faces significant challenges. Epistemic probability and causality can help out here: Epistemological. Objective Bayesian methods can be applied to evidence integration. Objective Bayesian nets model dependencies. Recursive Bayesian nets also model causal and mechanistic structure. Metaphysical. Epistemic causality can account for multifarious evidence.
Bibliography Galas, D. J. and Hood, L. (2009). Systems biology and emerging technologies will catalyze the transition from reactive medicine to predictive, personalized, preventive and participatory (P4) medicine. Interdisciplinary Bio Central, 1(6):1 5. Landes, J. and Williamson, J. (2013). Objective Bayesianism and the maximum entropy principle. Entropy, 15(9):3528 3591. Vandamme, D., Fitzmaurice, W., Kholodenko, B., and Kolch, W. (2013). Systems medicine: helping us understand the complexity of disease. Quarterly Journal of Medicine, 106(10):891 895. Williamson, J. (2005). Bayesian nets and causality: philosophical and computational foundations. Oxford University Press, Oxford. Williamson, J. (2010). In defence of objective Bayesianism. Oxford University Press, Oxford.