Digital Health Lab Day – Wädenswil – 03.10.2019 Sven Hirsch, Norman Juchler Real and assumed insights – statistical models and imaging biomarkers for disease characterization of intracranial aneurysms
Clinical data is created primarily to treat patients Structured Primarily driven by Primarily clinical needs research-driven database Data Analysis / acquisition modelling Tools 80% 2
Working with clinical data is a challenge! Challenges § Access procedures (ethical approval, anonymization) § Missing links between different sources § Data often lacks harmonization and documentation Solutions § Data tends to be inconsistent § Data skewed by selection bias § Curation of consistent research databases § Robust pipelines for data processing § Transfer of clinical domain knowledge required § Always consider selection bias 3
Machine learning for disease characterization 5
Intracranial aneurysms are focal deformations of cerebral arteries § Typically located in the direct proximity of vessel bifurcations and the Circle of Willis § 3% of the population is affected 1 § Mostly free of symptoms – until rupture § Rupture risk ~1% per year 2 § Mortality 15%, invalidity 31% upon rupture X-ray Angiography Illustration of cerebral vasculature Source: AneuX, HUG-p163. with an intracranial aneurysm. Image size: 256x256px, (Source: Sentera Healthcare) Voxel size: ~ 0.25mm 1 M. Vlak et al. "Prevalence of unruptured intracranial aneurysms”, 2011 2 A. Ahmed et al. “Aneurysms, intracranial”. Encyclopedia of the Neurological Sciences. 2014 6
From shape to prognosis… Vision: Shape AneuX AneurysmDatabase as bio-marker for Data: Medical disease progression § 1350+ patient records imaging data and clinical data § 900+ medical images A multicentric initiative to improve the treatment of § 350+ aneurysm geometries intracranial aneurysms aneurIST database … Morphology study Irregularity Evaluation of Rating scoring Irregularity Study on study schemes location Rating dependency study 8
Quantitative morphology for disease status prediction § Size and shape are associated with pathologic wall conditions § Irregular shape or “ugliness” already used in clinics as subjective risk indicator § How to quantify irregularity/morphology? Classical approach: Rupture: Yes / No? 9
Quantitative morphology for disease status prediction Shape Curvature Writhe-number Zernike Moment Size based descriptors Invariants § Aspect ratio Various curvature § Volume “energies” Captures surface Generic shape § Non-sphericity § Surface area asymmetries and descriptor § Ellipticity § Aneurysm size “geometric § ZMI cumulants § Undulation § Height deformation” § Designed for § Bottleneck factor § Max. diameter energies shape queries § … § Neck diameter § … 10
Result: Aneurysm morphology carries significant information about the disease status Prediction performance for exemplary descriptors Feature configurations: Prediction based on 11. Normalized Gauss Prediction based multiple indices on single indices 1. Volume (V) 12. Size: 2. Surface (S) Size 13. Size + AR: 3. Neck diameter (Dn) NSI Shape 14. Size + AR: Size All 4. Dome height (H) 15. Shape: 5. Aneurysm size (ASZ) 6. Non-sphericity index (NSI) 16. Shape: 7. Ellipticity index (EI) 17. Curvature: Shape Curvature 8. Undulation index (UI) 18. Shape + size: 9. Aspect ratio (AR) 19. Size: 10. Normalized mean curvature 20. All : 11. Normalized Gaussian curvature Classifier: SVM, trained with 6-fold nested cross- validation and 50 repetitions, using 450 datasets 11
Result: Aneurysm morphology carries some information about the disease status …but the initial expectations were not met Moderate overall prediction accuracy (0.75 in the best case) § Morphometric predictors possibly not specific enough? § Labelling problem? (Rupture status vs. stability status) § Selection bias? § Missing clinical data for stratified analysis § How to integrate tools into clinical routine? § 12
We had to revise our engineering approach… Improvements: Vision: Shape as bio-marker for § Curation of a complete, validated database Data: Medical disease progression imaging data and § Realignment with clinical needs clinical data § Follow previously defined clinical hypotheses § Move away from rupture status prediction Morphology study Irregularity Evaluation of Rating scoring Irregularity Study on study schemes location Rating dependency study 13
Integration of clinical knowledge 14
Irregularity: A “crowd-sourced” metric for shape § What constitutes an “irregular” shape? § Which morphometrics represent perceived irregularity the best? § What is the relationship between irregularity and clinical factors? 15
We collected qualitative rating data using our interactive rating tool § Focus on subjective assessment of geometries § Aggregation using rank-based analysis Irregular Regular 9 1 16
Perceived irregularity is associated with rupture § Critical characteristics: 1.0 ICA oSh 1.0 Rupture Sex Smoking Hypertension IA size (aSz) Patient age 0CA 01 78 unruptured 36 male 40 non-smokers 70 non-hypertens. 134 aneurysms 104 aneurysms Presence of blebs/lobules § 41 ruptured 98 female 74 smokers 49 hypertensive Mean: 6.9mm Mean: 53.0y PCoPA Std: 3.4mm 12.1y 119 total 134 total 114 total 119 total Std: ACoPA 0.8 0.8 Characteristic AUC p -val AUC p -val AUC p -val AUC p -val Sp p -val Sp p -val Asymmetric aspect § Rough surface 0.59 0.12 0.55 0.41 0.51 0.90 0.50 0.97 0.68 *** 0.03 0.76 TPR (true positive rate) Elongation / non-sphericity § Blebs 0.71 ** 0.54 0.53 0.56 0.32 0.52 0.70 0.50 *** 0.10 0.33 0.6 0.6 irregularity irregularity (AUC=0.81) Lobules 0.79 *** 0.64 0.01 0.52 0.68 0.52 0.69 0.40 ** 0.02 0.83 asymmetry (AUC=0.81) Curvature and total writhe § Asymmetry 0.81 *** 0.59 0.10 0.51 0.85 0.53 0.62 0.49 *** 0.09 0.39 lobules (AUC=0.79) blebs (AUC=0.71) Complex vasc. 0.51 0.85 0.56 0.28 0.52 0.76 0.57 0.17 0.13 0.28 0.03 0.76 § No association was found 0.4 0.4 rough surface (AUC=0.59) Irregularity 0.81 *** 0.58 0.14 0.50 0.90 0.51 0.79 0.71 *** 0.08 0.42 with other clinical factors… complex vasc. (AUC=0.51) Aneurysm size 0.71 ** 0.62 0.03 0.57 0.18 NSI (AUC=0.83) 0.52 0.24 - - 0.14 0.16 0.2 0.2 GLN (AUC=0.78) Non-sphericity 0.83 *** 0.60 0.07 0.55 0.38 0.50 0.97 0.61 *** 0.01 0.93 aSz (AUC=0.71) Curvature 0.78 *** 0.58 0.15 0.55 0.34 0.52 0.78 0.81 *** 0.09 0.38 irregularity + loc. (AUC=0.87) § …except for aneurysm 0.0 Patient age 0.58 0.21 0.59 0.18 0.55 0.48 0.74 * 0.14 0.16 - - 0.0 unruStured ruStured 0.0 0.2 0.4 0.6 0.8 1.0 location ruSture6tatus FPR (false positive rate) § Metrics for aneurysm Irregularity is an independent risk factor, morphology shows strong along with location (and aneurysm size ) dependency on location 17
The power of bigger data 20
Data infrastructure and harmonization An ongoing effort aneurIST ISGC AneuX towards a multi- 2006 - 2010 2014 - … 2006 - … centric, research- oriented data collection 21
Tool for case matching and statistical analysis § Select cases according to predefined criteria § Useful to compare subcohorts of patients and to “skim” the data for statistical relationships § Intuitive representation of simplifies the interpretation of results 22
Example: Mosaic plots to compare multiple and solitary aneurysms Independent variable (predictor) Location of ruptured or larger IA Dependent variable (response) AComA Solitary IA V-B Multiple IAs Pcom MCA § AComA most common location in solitary IA § Useful to compare two categorical variables § MCA most common location in multiple IAs § Width of cols and rows indicates relative proportions § Female have more multiple IAs § Color encodes the outcome of statistical tests: § Smokers more likely having multi IA Blue: significantly over-represented § Red: significantly under-represented § § Multiple IAs rupture less 23
Application: Assess clinical tools (Example: PHASES score) 24
Conclusions / Experiences § Always consider biases! § Visualization, statistical tests and perseverance are your best friends § The AneuX AneurysmDataBase is one of the largest databases of its kind § Beware of clinical scores based on univariate analysis § Better tools are required to accommodate the multifactorial nature of the disease (e.g. Bayesian network reasoning) 25
Acknowledgments Institute of Applied Simulation , ZHAW University Hospitals Geneva Norman Juchler Philippe Bijlenga § § Sabine Schilling Sandrine Morel § § Erich Zbinden Nicolas Roduit § § Nicolas Dupuy § Rafik Ouared § Vital-IT , Lausanne Hirslanden Klinik , Zurich Vital-IT Jérôme Dauvillier Daniel Rüfenacht § § Robin Liechti Isabel Wanke § § Olivier Martin Stephan Wetzel § § 26
Recommend
More recommend
