1 / 38

Algorithms and tools to support medical decision making -Case studies-

This research study explores the use of algorithms and tools in medical decision making through case studies. It covers topics such as cardiovascular disease diagnosis, remote monitoring, telemedicine, user need elicitation, and health technology assessment.

pdon
Télécharger la présentation

Algorithms and tools to support medical decision making -Case studies-

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Algorithms and tools to support medical decision making -Case studies- Leandro Pecchia

  2. Me RESEARCH BACKGROUND Ω≈f(stress, fat, salary, ↓free-t,…) • Careerpath: 2013-2016: Assistant Professor, University of Warwick 2011-2013: Research Fellow (RF2), University of Nottingham 2009-2011: Research Fellow (RF1), UNINA* 2005-2009: PhD in Biomedical Engineering, UNINA* May 2005: BSc+MSc in Electronic Eng., UNINA* *UNINA= University Federico II of Naples, Italy PHD RF1 UNINA RF2 NOTT Ass. Prof. 2005 2009 2011 2013 2007 • Research interests: • Biomedical signal processing and second level pattern recognition/data-mining • Early stage Health Technology Assessment (HTA) and User Need Elicitation methods • My main applications: • active/healthy ageing: chronic cardiovascular diseases and falls in elderly • Disease Management Programs, patient pervasive monitoring and Telemedicine

  3. Me Contributions and outputs (1/2) • Signal processing and pattern recognition for Cardiovascular disease (CVD) • to identify Congestive Heart Failure (CHF) [early diagnosis] • 2011, “Long-term HRV & CHF detection ”, Med & BiolEngin & Computing, 49 (1):67- 74 • 2011, “Short-term HRV & CHF detection”, IEEE T InfTechnolog in Biomed, 15 (1):40-46. • to manage chronic CVD monitoring its damages and severity… [early detection of risks] • 2012, “HRV &Organ Damage in Hypertension”, BMC Cardiovascular Disorders, 12:105 • 2013, “Long-term HRV & CHF severity assessment”, IEEE J. of Biom. and Health Informatics, 17(3): 727-733 • …also in remote monitoring applications: [telemedicine] • 2011, “Remote Health Monitoring of CHF”, IEEE T Bio-Med Eng, 58 (3):800-804 • 2011, “A feasibility study on telemedicine”, Biomedical Engineering Online, 10: 49 • Other signal processing and pattern recognition applications • 2011, “Nonlinear HRV for real-life stress detection”, Biomedical Engineering Online 10: 96 • 2012, “Pupillometric analysis for assessment of gene therapy”, Biomedical Engineering Online,11(1):40 • 2013, “Infant cry analysis for early detection of Autism”, ICHI2013 (+ submitting 2 journal papers)

  4. Me Contributions and outputs (2/2) • Medical decision making is complex and multidisciplinary and needs: • Quantitative knowledge: from the best available evidence (RCT, meat-analyses, network meta-analyses) • Qualitative knowledge: to interpret the top of the EBM pyramid into everyday clinical practice • Methods for the “impact” of BME researches: quantify qualitative knowledge • User need elicitation using the Analytic Hierarchy Process (AHP) method • 2013, “User needs elicitation via AHP”, BMC Medical Informatics and Decision Making, 3(1):2 • 2013, “AHP & auto-injection of epinephrine”, HIS2013, 25-27 March 2013 in London. • 2011,“Factors affecting wellbeing in elderly”, ISAHP 2011, Sorrento, Naples, Italy. • Health Technology Assessment (HTA), especially for early stages of technology development • 2013, “HTA & AHP”. In Studies in Fuzziness and Soft Computing, ed. Springer, Volume 305, • 2013, “ HTA, Telemedicine, CHF”. In Telehealthcare Computing and Engineering: Principles and Design. Science Publishers. ISBN 978-1-57808-802-7 (book chapter) • 2013. “Enhanced Remote Health Monitoring.. In Telehealthcare Computing and Engineering: Principles and Design, ed. Science Publishers. ISBN 978-1-57808-802-7 (book chapter) • 2012, “Network meta-analysis & mini-invasive surgery”, Surgical Endoscopy,2012 Jun 16, [Epub ahead of print] • 2012, “RCT for innovative biological drug”, Hernia, 20 November 2012 Nov, [Epub ahead of print] • 2011, “Meta-analysis & minimally invasive surgery”, Minimally Invasive Therapy & Allied Technologies, 21(3):150-60 • June 2012, Treasurer of the HTA Division of International Federation of BME, IFMBE • Risk factors for falls in elderly home dwelling • Many intrinsic risk factors are related to physiological condition that can be detected • 2011, “Risk factors for falls”, Methods of Information in Medicine, 50 (5):435 • 2010, “Risk factors for falls”, International Journal of the Analytic Hierarchy Process, 2 (2)

  5. -Case study 1- SHARE Project*: home monitoring for patients suffering from Congestive Heart Failure (CHF) Leandro Pecchia *Smart Health and Artificial intelligence for Risk Estimation grant PON04a3_00139 to PM; 2007-2013 Italian National Operational Programme for Research and Competitiveness. PI: Dr Paolo Melillo

  6. INTRO INTRO CS1 • Congestive Heart Failure (CHF) is • leading cause of hospitalization among the elderly in developed countries • up to 50% of patients are rehospitalized within 3 months. • Mortality ranges from 10% in patients with mild HF to 40% in severe cases • COSTS: direct treatment costsof HF represent 2–3% of the total healthcare budget • In literature there are three main models of care: • Usual care (UC): outpatient follow-up (GP guided) as recommended by guideline • Disease Man. Programs (DMP): UC + specialized doctors/nurses proactively at home • Home Monitoring (HM): DMP + Information and Communication Technologies (ICT) • The goal of HM, DMP and UC for CHF is to: • ↑QoL (or at least maintain stable) • ↓mortality (for CHF and for all causes), • ↓NHS costs by: ↓ bed days (CHF/all causes); ↓readmissions (CHF/all causes)

  7. GOALS GOALS CS1 • Not all the HM is equally important and independent contribution of ICT is unclear • Thus, the goals of the SHARE project are: • To identify those elements that make HM more effective than DMP and UC (WP1) • To design the most effective HM program • To develop the ICT system to support such a HM • To test its cost-effectiveness with 2 clinical trials WP1: HTA of HM WP2: HM and ICT platform R&D WP3: Algorithms R&D WP4: Observational Study WP5: Prospective study WP6: PM/Dissemination

  8. METHODS METHODS 1/2 CS1 • Meta-analyses were performed comparing DMPvsUC & HMvsUC • Only well designed RCT were included • Outcome considered: • ↑QoL (or at least maintain stable) • ↓mortality (for CHF and for all causes), • ↓ bed days (CHF/all causes); • ↓readmissions (CHF/all causes) • Classified all the HM RCT according to: patients’ severity and HM complexity • We studied the correlations between the last 6 outcomes and • patients’ severity • HM complexity • According to these, the clinical trial and the ICT platform were designed • Algorithms for early detection of CHF worsening were • Developed using public DB available • Adaptedusing the information collected during the WP4 [now] • Integrated into the ICT platform [2014] • Assessment of cost-effectiveness in WP5 [2016]

  9. METHODS METHODS 2/2 CS1 • HM RCT were classified according to patients’ severity and HM complexity • Correlations effectiveness-severity and effectiveness-complexity were computed MEAN PATIENTS AGE DATA MONITORED HM COMPLEXITY Signal & Parameters & symptoms 75 - PATIENT SEVERITY Parameters & symptoms 70 - Only Symptoms Weekly NYHA II -- Daily CNYHA III -- - - NYHA CLASSES 40 - FREQUECY OF THE MONITORING 35 - EJECTION FRACTION

  10. RESULTS RESULTS CS1 • Study selection: 314 papers Title and abstract seletion: 8 Not heart failure 18 Invasive hemodynamic monitoring 43 Editorial or review 70 not an RCT 41 Other heart failure intervention or research 19 Study design Full paper selection: 1 Not heart failure 7 Editorial or review 16 not an RCT 40 Other heart failure intervention or research 10 Study design 4 Other languages 5Short Follow-up 115 papers 32 papers 23 DMP vs UC 8 HM vs UC 1DMP & HM vs UC

  11. RESULTS RESULTS CS1 • There is evidence that DMP are more effective that UC • Reducing All-causes mortality • Reducing readmission (All-causes and HF-related) • It seems, but there is not evidence that DMPs reduce bed-days

  12. RESULTS RESULTS CS1 • Mortality: HM vs UC • All causes HF mortality The HM RCT seems (no statistically significant result) more effective than UC

  13. RESULTS RESULTS CS1 • Readmission: HM vs UC • All causes HF mortality The HM RCT seems (no statistically significant result) more effective than UC

  14. RESULTS RESULTS CS1 • Bad days: HM vs UC • All causes HF mortality The HM RCT seems (no statistically significant result) more effective than UC

  15. RESULTS RESULTS CS1 • HM RCT classification: Patients’ severity

  16. RESULTS RESULTS CS1 • HM RCT classification: HM protocol complexity

  17. RESULTS RESULTS CS1 • HM outcomes vs Pz complexity and HM complexity: Survival Saved Bed days Reospedalization significant correlation No significant correlation No significant correlation PATIENT SEVERITY HM COMPLEXITY significant correlation significant correlation No significant correlation • Evidence: • HM is more effective for more severe patients (SHARE now focus on NYHA >2) • More complex HM are more effective (SARE is designed accordingly) 17/32

  18. RESULTS RESULTS CS1 • How this informed the SHARE the clinical protocol? • Enrolling a proper number of severe cases [about 300 subjects in 12 months] • Acquiring daily useful symptoms, parameters and signals • These info will be daily reviewed to early detect patients’ worsening • …and how the ICT platform? • The DSS (WP3) will support clinician in modulating patient therapy • What these algorithms does and how they look like? BioHarmess3, Zephyr • sensors: ECG, 3axial accellerations, Breathing • up to 3 days recording memory • Communication: Bluetooth/ZigBee • Memory: Up to 7 days recording memory • Cost: £350 (≤10) or £200 (>10) NEXUS10(4), MindMedia • sensors: • EXG: ECG, EMG, EOG, EEG • SpO2, BVP, Body Temperature, Breathing acts, GSR • Communication: Bluetooth • Memory: Up to 7 days recording memory • Costs: from £5k to £12k, according to the sensors Healthcare Professionals (App) BIOMEDICAL SIGNAL PROCESSING (wearable devices) REMOTE PROCESSING (DDS) WARNING APPROPRIATE INTERVENTION 18/32

  19. RESULTS RESULTS CS1 • Autonomous Nervous System (ANS) controls human equilibrium (homeostasis) • Normal subjects show a good degree of variability in body functionalities, reflecting a continuous state of unstable equilibrium • Unstable equilibrium is complex to control, but allows faster state changes • This allow humans to react promptly to: • internal changes (i.e. emotions, stress) • external treads (i.e. the lion…) • Monitoring these changes we can estimate the status of a subject and how stable it is… 19/32

  20. RESULTS RESULTS CS1 • Peculiarities of these signals: • Bandwidth (0- few tens of Hz) • Low S/N ratio in band • No stationary (FFT cannot be used!) • Strong non linearity and high dependence from parameters (chaos?) • Problems for pattern recognition • Limited cases • Natural patterns (no human-generated) • Last but not least… Signal pre-processing: filtering, beat recognition (normal vs abnormal),… Signal processing: features extraction in time-, frequency-, nonlinear- domain Pattern recognition: signal/patient classifications “normal vs CHF”, “mild vs severe”, “damage vs sane” …our methods/results are needed by clinicians that cold be not skilled in mathematical methods! 20/32

  21. RESULTS RESULTS CS1 Detection: long & short HRV (Guidelines says that 12-leads ECG is not enough to diagnoses HF) Severity (NYHA) assessment ECG PREPROCESSING C) B) A) HRV DETECTION HRV FEATURES EXTRAC. FEATURES COMBINATION CART TRAIN/TEST PERFORMANCE ASSESSMENT • HRV to identify Congestive Heart Failure (CHF) • A) 2011, Med & BiolEngin & Computing 49 (1):67- 74 • B) 2011, IEEE T InfTechnolog in Biomed 15 (1):40-46. • HRV to manage CHF monitoring its severity… • C) IEEE J. of Biom. and Health Informatics, 17(3): 727-733 • D) BMC Cardiovascular Disorders, 12:105 [organ damages] 21/32

  22. CONCLUSIONS CONCLUSIONS • Not all the HM strategies are equally effective: • HM is more effective on more severe patients • more complex HM interventions seems more effective than less complex ones • However, the increased quantity of information requires: • Reliable technological solution • Smart algorithms to extract the useful information • Integrated management strategies • The preliminary results of the algorithms developed are promising on public DB • These SHARE trials will generate reliable databases for the adaptation of these algorithms.

  23. -Case Study 2- A software tool to support the Health Technology Assessment (HTA) and the user need elicitation of medical devices via the Analytic Hierarchy Process (AHP) L. Pecchia1, F. Crispino2, S. Morgna3 1University of Warwick, The United Kingdom 2 Business Engineering, Avellino, Italy 3 University of Nottingham, The United Kingdom l.pecchia@warwick.ac.uk

  24. INTRODUCTION METHOD RESUTLS COMCLUSIONS HTA/UNE AHP for HTA & User Need Elic. CS2 How to prioritize the needs? NEED ANALYSIS INDIVIDUATION CLASSIFICATION PRIORITIZATION IDENTIFY EXISTING TECHNOLOGIES How to measure the MD performance in non-clinical domains? MULTIDIMENSIONAL EVALUATION … ECONOMICAL ETHIC/ SOCIAL CLINICAL EPIDEMIOLOGICAL DATA ANALYSIS How to measure the fitting between MD performance and needs?? RELATIVE ASSESSMENT EFFICACY EFFICIENCY PERFORMANCE 24/32

  25. INTRODUCTIONMETHOD RESUTLS COMCLUSIONS AHP AHP for HTA Hierarchy via an exempla CS2 Developing (or selecting) the health technology for a clinical problem (i.e. congestive heart failure) Technological domain (services/spare parts/ Human F) [Medical Eng.] Clinical domain (effectiveness/utility) [clinicians/cardiologists/ger.] Economical domain (costs) [Hosp. Managers] usability education service … ↓mortality ↓ worsening … ↑ qaly Initial cost ReadmissionC. … ALTERNATIVE 1 Disease Management Program ALTERNATIVE 2 Telemedicine ALTERNATIVE 3 Active Implantable Device • How important is each need for the assessment? [needs prioritization] • How each alternative satisfy each factor? [MD performance] • How each alternative fit with the goal? [MD/Goal fitting]

  26. INTRODUCTIONMETHOD RESUTLS COMCLUSIONS AHP Numerical values Much more More important Equally important Less important Much less important AHP method pairwise comparisons Process CS2 NEEDS’ INDIVIDUATION EXPERTS (5) (3) TREE OF NEEDS (1) (1/3) QUESTIONNAIRES (1/5) JUDGEMENTS MATRIX (J) CONSISTENCY RATIO(CR) Eigenvector (priorities) Eigen value (coherence) CR >0.1 IF DATA POOLING RELATIVE IMPORTANCE RELATIVE IMPORTANCE OF NEEDS OF NEEDS’ CATEGORIES N1>> N3 ALTERNATIVES’ N1 > N3 N1>N2 & N2>N3 => PERFORMANCE ASSESSMENT N1< N3 ALTERNATIVES’ PRIORITIZATION

  27. INTRODUCTIONMETHOD RESUTLS COMCLUSIONS AHP method Analytic needs prioritization CS2 Method Developing (or selecting) the health technology for a clinical problem (i.e. congestive heart failure) Technological domain (services/spare parts/ Human F) [Medical Eng.] Clinical domain (effectiveness/utility) [clinicians/cardiologists/ger.] Economical domain (costs) [Hosp. Managers] usability education service ↓worsening ↓ mortality ↑ qaly Initial cost ReadmissionC. ALTERNATIVE 1 Disease Management Program ALTERNATIVE 2 Telemedicine ALTERNATIVE 3 Active Implantable Device

  28. INTRODUCTIONMETHOD RESUTLS COMCLUSIONS AHP AHP for HTA CS2 Global importance ALTERNATIVE 1 DMP ALTERNATIVE 2 Telemedicine ALTERNATIVE 3 Active Implantable D. usability education service ↓worsening ↓ mortality ↑ qaly Initial cost ReadmissionC. 28/32

  29. INTRODUCTIONMETHODRESUTLS COMCLUSIONS AHP The system a web tool with App CS2 http://www.ahpapp.net/ • Method: AHP for HTA/User need elicitation • Applications: Publication in Healthcare whit the App • Models: to be downloaded and adapted in your study • Community: experts willing be involved 29/32

  30. INTRODUCTIONMETHODRESUTLS COMCLUSIONS AHP AHP for HTA Hierarchy via an exempla CS2 • Users: • The elicitor: • design/pilot the hierarchy and the questionnaires, • invites domain experts and final responders, • pool the results; • generate the report; • publish the results on the web portal. • The domain expert: • review the hierarchy/questionnaires, • suggest final responders or other domain experts; • The final responder: • under invitation, • download the hierarchy • answer the questions.

  31. INTRODUCTIONMETHODRESUTLS COMCLUSIONS AHP AHP for HTA Hierarchy via an exempla CS2 • Two possible scenarios: • S1: Local • elicitor, domain experts and the final responders in the same place • Using the APP to speed-up the process and find consensus • S2:Remote • elicitor, domain experts and the final responders NOT in the same place, • Using the APP and the portal to cooperate to the study via the web. • Functionalities: • Create the Hierarchy: problem definition/hierarchy draft • Download an existing hierarchy: to be used as starting model • (only S2) Invite domain experts: study piloting • (only S2) Amend the hierarchy • (only S2) Invite responders • (only S2) Participate to the study • Analyse and Pool results • Generate a report • Publish: upload on the portal hierarchy ¦¦ results ¦¦ reports¦¦papers

  32. INTRODUCTION METHOD RESUTLS COMCLUSIONS Concluding CONCLUSIONS • This is the first tool specifically designed to: • perform shared decision making in healthcare • involve lay-users into the decisional process (paramount important for HTA) • applying the AHP to the HTA/the user need elicitation in healthcare • Medical Decision making is complex (…not necessary difficult!) • Methods have to be: • Reliable • Well tested according to clinical practices • Intelligible (no black boxes) • Easy to use/understand for people not skilled in maths • Traceable (you may have to prove that you did the best you could after years) 32/32

  33. Thank you! Leandro

  34. HTA NEED ANALYSIS INDIVIDUATION CLASSIFICATION SCORING IDENTIFY EXISTING TECHNOLOGIES MULTIDIMENSIONAL EVALUATION … ECONOMICAL ETHIC/ SOCIAL CLINICAL EPIDEMIOLOGICAL DATA ANALYSIS RELATIVE ASSESSMENT (VERSUS BENCHMARK) EFFICACY EFFICIENCY PERFORMANCE reduce price Reduce data uncertainty DISSEMINATION OF INFORMATION MONITORING HTA standard methods COST DIFFERENCE + * - + - EFFECT DIFFERENCE (QALY?) * NHS NICE willingness-to-pay ‘threshold’ range of £20,000-£30,000 per QALY. Example 4 – price optimisation Example 5 – highlighting data requirements Example 3 – not cost effective Example 4 – price optimisation Example 1 – cost effective 34/33

  35. HTA limits HTA Limits of standard methods VS L Pecchia, MP Craven, “Early stage Health Technology Assessment (HTA) of biomedical devices. The MATCH experience”. World Congress on Medical Physics and Biomedical 2012, 26-31 May 2012, Beijing, China. 35/33

  36. Method 1 Headroom Analysis p’2 DC p2 HEADROOM OTHER NHS Costs p3 DEVICE Production Costs 30k£/QALY p1 DU DQALY http://www.match.ac.uk/ 36/33

  37. MM 1-pd-pw-pe 1-pd-p’w-p’e 1-pwb-pwd-p’we 1-pwb-pwd-pwe 1-peb-pew-ped 1-peb-pew-ped 4 states Markov Models disease/worsening/exacerbation/dead HAVE DISEASE (C1,U1) DEAD (C2,U2) HAVE DISEASE (C’1,U1’) DEAD (C2’,U2’) INNOVATE DEVICE pd pd peb peb pe p’e pwb pw pwb p’w ped ped pwd WORSE DISEASE (C3,U3) pwd WORSE DISEASE (C’3,U3’) pwe pwe EXACERBATION (C4,U4) EXACERBATION (C4’,U4’) pwe P’we ↓worsening (p’w < pw) DC ↓exacerbation (p’e< pe &p’we < pwe) DC/DU =30K/QALY DU http://www.match.ac.uk/ 37/33

  38. MM&AHP 1-pd-pw-pe 1-pd-pw-pe 1-pwb-pwd-pwe 1-pwb-pwd-pwe 1-peb-pew-ped 1-peb-pew-ped Markov Models & AHP What if some information are missing (eHTA)? Missing data can be estimated using AHP... HAVE DISEASE (C1,U1) DEAD (C2,U2) HAVE DISEASE (C’1,U1’) DEAD (C2’,U2’) INNOVATE DEVICE pd ± Dp pd peb peb pe pe pwb pw pwb pw ped ped pwd WORSE DISEASE (C3,U3) pwd WORSE DISEASE (C’3,U3’) pwe pwe EXACERBATION (C4,U4) EXACERBATION (C4’,U4’) pwe pwe …and using sensitivity analysisto estimate worst/best cases. 38/33

More Related