phonemd
play

PhoneMD: Learning to Diagnose Parkinsons Disease from Smartphone - PowerPoint PPT Presentation

Aug 22, 2023 •104 likes •615 views

PhoneMD: Learning to Diagnose Parkinsons Disease from Smartphone Data Patrick Schwab and Walter Karlen @schwabpa @mhsl_ethz Institute of Robotics and Intelligent Systems ETH Zurich Parkinsons Disease (PD) Slow degeneration of motor

  1. PhoneMD: Learning to Diagnose Parkinson’s Disease from Smartphone Data Patrick Schwab and Walter Karlen @schwabpa @mhsl_ethz Institute of Robotics and Intelligent Systems ETH Zurich

  4. Parkinson’s Disease (PD) • Slow degeneration of motor skills • Hard to diagnose • Assessment of symptoms • Similar symptoms in other diseases • Symptom fl uctuations • Only ~80% of diagnoses are accurate 1 • ~7m (0.3%) a ff ected, 120,000 deaths 2 1 Rizzo, G. et al. (2016) Accuracy of clinical diagnosis of Parkinson disease: A systematic review and meta-analysis. Neurology 86 (6). 2 de Lau, LM and Breteler MM. (2006) Epidemiology of Parkinson's disease. Lancet Neurology 5 (6). 4

  5. Wide Variety of Symptoms Cognition Speech Dexterity Movement 5

  6. Motor Impairments Rigidity Tilted Posture Reduced Arm Movement Tremor of Extremities Shu ffl ing Gait & Short Steps 6

  7. The Idea Can we use machine learning on long-term smartphone data to diagnose Parkinson’s? Likelihood of PD 7

  8. The Dataset 8

  9. The mPower Study • We use data collected in the mPower study 1 • Openly available on Synapse 2 • App users (with and without Parkinson’s, n=1853) were asked to perform several tests regularly • Outcome: Prior clinical PD diagnosis 1 Bot, B.M., et al. (2016) The mPower study, Parkinson disease mobile data collected using ResearchKit. Scienti fi c data 3. 2 Synapse Platform, https://www.synapse.org/#!Synapse:syn8717496 (Accessed: Nov 13, 2017) 9

  10. Tests Overview � � � � � � walking tapping memory voice � � � � � � � 1 � ? tests 2 ? ○ ○ x signals � � �� ○ ○ ○ y � z 10

  11. Tests Overview � � � � � � walking tapping memory voice � � � � � � � 1 � ? tests 2 ? ○ ○ x signals � � �� ○ ○ ○ y � z 11

  12. mPower Walking Test • In the walking task, participants were asked to do the following three-segment task (each ~30s): 12

  13. mPower Walking Test • In the walking task, participants were asked to do the following three-segment task (each ~30s): 13

  14. mPower Walking Test • In the walking task, participants were asked to do the following three-segment task (each ~30s): (1) Walk outbound 14

  15. mPower Walking Test • In the walking task, participants were asked to do the following three-segment task (each ~30s): (2) Rest (1) Walk outbound 15

  16. mPower Walking Test • In the walking task, participants were asked to do the following three-segment task (each ~30s): (2) Rest (1) Walk outbound (3) Walk return (incl. turn) 16

  17. Data Streams • Accelerometer time series: • Acceleration • Rotation Rate • Attitude 17

  18. Tests Overview � � � � � � walking tapping memory voice � � � � � � � 1 � ? tests 2 ? ○ ○ x signals � � �� ○ ○ ○ y � z 18

  19. mPower Voice Test � “aaaaaah” � 19

  20. Data Streams • Voice recording • 44100 Hz • ~30 seconds 20

  21. Tests Overview � � � � � � walking tapping memory voice � � � � � � � 1 � ? tests 2 ? ○ ○ x signals � � �� ○ ○ ○ y � z 21

  22. mPower Tapping Test 1 Bot, B.M., et al. (2016) The mPower study, Parkinson disease mobile data collected using ResearchKit. Scienti fi c data 3. 22

  23. Tests Overview � � � � � � walking tapping memory voice � � � � � � � 1 � ? tests 2 ? ○ ○ x signals � � �� ○ ○ ○ y � z 23

  24. Tests Overview 1 Bot, B.M., et al. (2016) The mPower study, Parkinson disease mobile data collected using ResearchKit. Scienti fi c data 3. 24

  25. Approach 25

  26. 1- Hierarchical Approach Per-test Models: Specialised in each test type. model input output P � x � y � � x � y � P ○ ○ x � y � � ○ ○ ○ P � � � �� x � � y � P 26

  27. 1- Hierarchical Approach Per-test Models: Specialised in each test type. model input output P � x � y � � x � y � P ○ ○ x � y � � ○ ○ ○ P � � � �� x � � y � P Independent models 27

  28. 2- Hierarchical Approach Evidence Aggregation Model (EAM): Integrate available test data over time. ( m � ,1 , y � ,1 ) ( m � ,2 , y � ,2 ) ( m � ,3 , y � ,3 ) ( m � ,4 , y � ,4 ) EAM h 2 h 3 h 4 h 1 y 28

  29. 2- Hierarchical Approach Evidence Aggregation Model (EAM): Integrate available test data over time. Any number of tests ( m � ,1 , y � ,1 ) ( m � ,2 , y � ,2 ) ( m � ,3 , y � ,3 ) ( m � ,4 , y � ,4 ) EAM h 2 h 3 h 4 h 1 y 29

  30. 2- Hierarchical Approach Evidence Aggregation Model (EAM): Integrate available test data over time. Any number of tests ( m � ,1 , y � ,1 ) ( m � ,2 , y � ,2 ) ( m � ,3 , y � ,3 ) ( m � ,4 , y � ,4 ) EAM h 2 h 3 h 4 h 1 y Final diagnostic score 30

  31. Neural Soft Attention 1 allows us to relate the decisions • A soft attention mechanism 2 and (2) tasks. to the most relevant (1) input segments Other vs. all Other: 67 % (s) 1 Bahdanau, D. et al. (2014). Neural Machine Translation by Jointly Learning to Align and Translate. ICLR. 2 Schwab, P., et al. (2017). Beat by Beat: Classifying Cardiac Arrhythmias with Recurrent Neural Networks. Computing in Cardiology. 31

  32. Results & Discussion 32

  33. Results on Test Set 0,00 0,25 0,50 0,75 1,00 EAM (Both) + age + gender EAM (Neural networks) + age + gender Predictive Performance [AUC] EAM (Feature) + age + gender EAM (Both) EAM (Neural networks) EAM (Feature) Mean Aggregation (Neural networks) Mean Aggregation (Feature)

  34. Results on Test Set 0,00 0,25 0,50 0,75 1,00 EAM (Both) + age + gender Signi fi cantly better than demographic baseline EAM (Neural networks) + age + gender Predictive Performance [AUC] EAM (Feature) + age + gender EAM (Both) EAM (Neural networks) EAM (Feature) Mean Aggregation (Neural networks) age + gender Mean Aggregation (Feature)

  35. Results on Test Set 0,00 0,25 0,50 0,75 1,00 EAM (Both) + age + gender EAM better integrates the available test data EAM (Neural networks) + age + gender Predictive Performance [AUC] EAM (Feature) + age + gender EAM (Both) EAM (Neural networks) EAM (Feature) Mean Aggregation (Neural networks) Mean Aggregation (Feature)

  36. Results on Test Set 0,00 0,25 0,50 0,75 1,00 EAM (Both) + age + gender EAM (Neural networks) + age + gender Predictive Performance [AUC] EAM (Feature) + age + gender EAM (Both) Expert-designed and learned features comparable in EAM (Neural networks) performance EAM (Feature) Mean Aggregation (Neural networks) Mean Aggregation (Feature)

  37. Results on Test Set 0,00 0,25 0,50 0,75 1,00 EAM (Both) + age + gender EAM (Neural networks) + age + gender Predictive Performance [AUC] EAM (Feature) + age + gender Best model used both. EAM (Both) EAM (Neural networks) EAM (Feature) Mean Aggregation (Neural networks) Mean Aggregation (Feature)

  38. Neural Attention a test � � � � � 1 � � � � � � � � � � � � � 3 4 5 7 8 9 11 13 14 17 2 6 10 12 15 16 18 a seg a seg outbound rest 2 3 1 4 5 6 7 8 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

  39. Neural Attention Importance over tests a test � � � � � 1 � � � � � � � � � � � � � 3 4 5 7 8 9 11 13 14 17 2 6 10 12 15 16 18 a seg a seg outbound rest 2 3 1 4 5 6 7 8 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 Importance within test Importance within test

  40. Neural Attention (Subject with PD) a test � � � � � 1 � � � � � � � � � � � � � 3 4 5 7 8 9 11 13 14 17 2 6 10 12 15 16 18 a seg a seg outbound rest 2 3 1 4 5 6 7 8 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 Di ffi culty starting to move

  41. Neural Attention (Subject with PD) a test � � � � � 1 � � � � � � � � � � � � � 3 4 5 7 8 9 11 13 14 17 2 6 10 12 15 16 18 a seg a seg outbound rest 2 3 1 4 5 6 7 8 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 Abrupt stop

  42. Neural Attention (Subject with PD) a test � � � � � 1 � � � � � � � � � � � � � 3 4 5 7 8 9 11 13 14 17 2 6 10 12 15 16 18 a seg a seg outbound rest 2 3 1 4 5 6 7 8 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 Potential resting tremor

  43. Conclusion 43

  44. Conclusion • We present an approach to diagnosing PD that … ✔ • works based on multiple smartphone-based tests that cover a wide range of symptoms across long time frame ✔ • informs the clinician about the importance of tests and segments within those tests using neural attention ✔ • achieves strong performance in a representative cohort (n=1853) with an AUC of 0.85 (95% CI: 0.81, 0.89) • We highlight potential of smartphones as accessible tools for gathering clinically relevant data in the wild

Recommend

More recommend