# Multimodal Learning and Intelligent Prediction of Symptom Development in Individual Parkinson’s Patients

^{1}

^{2}

^{3}

^{4}

^{*}

## Abstract

**:**

## 1. Introduction

## 2. Materials and Methods

#### 2.1. Theoretical Basis

_{B}is the equivalence class of u, or a B-elementary granule. The family of all equivalence classes of IND(B) will be denoted U/I(B) or U/B. The block of the partition U/B containing u will be denoted by B(u).

**lower approximation**of symptoms set $X\subseteq U$ in relation to a symptom attribute $B$ as $BX=\left\{u\in U:{\left[u\right]}_{B}\subseteq X\right\}$, and the

**upper approximation**of $X$ as $\overline{B}X=\left\{u\in U:{\left[u\right]}_{B}\cap X\ne 0\right\}$. In other words, all symptoms are classified into two categories (sets). The lower movement approximation set $X$ has the property that all symptoms with certain attributes are part of $X$, and the upper movement approximation set has property that only some symptoms with attributes in $B$ are part of $X$ (for more details see [7]). The difference between the uppper and lower approximations of $X$ is defined as the boundary region of $X$: BN

_{B}($X$). If BN

_{B}($X$) is empty then $X$ is exact (crisp) with respect to B; otherwise if BN

_{B}($X$) $\ne \varphi $ then $X$ is not exact (i.e., is rough) with respect to B. We say that the B-lower approximation of a given set $X$ is union of all B-granules that are included in $X$, and the B-upper approximation of $X$ is of the union of all B-granules that have nonempty intersection with $X$.

**our rules must have some “flexibility”, or granularity**. In other words, one can also think about the probability of finding certain symptoms in a certain group of patients.

**This approach, called granular computation, should simulate the way in which experienced neurologists might interact with patients.**It is the ability to perceive a patient’s symptoms at various levels of granularity (i.e., abstraction) in order to abstract and consider only those symptoms that are universally significant and serve to determine a specific treatment(s). Looking into different levels of granularity determines the different levels of knowledge abstraction (generality), as well as assisting with greater understanding of the inherent knowledge structure. Granular computing simulates human intelligentce related to classification of objects in the sensory system (e.g., vision [2,3,4,5]) and motor [1] system as well as in task-specific problem solving behaviors.

#### 2.2. Computational Basis

#### 2.3. Dedicated Database

- Defining new attributes describing the Patient, Examination, and Measurement entities;
- Defining associated constraints or the dictionary of enumerative values for those attributes;
- Defining new types of files for storing additional information regarding Patients, Examinations, and Measurements;
- Establishing a numbering and naming convention for the instances of the above mentioned entities and files.

- Web service interfaces for contributing measurements made with mobile devices;
- Advanced privilege management functionality to allow multiple-team collaboration and support for different access levels;
- Internationalization in terms of user interface language and data and metadata language;
- Integration with the human motion database, to extend the scope of future examination.

## 3. Results

#### 3.1. Neurological Tests and Eye Movement Measurements

#### 3.2. Rough Set and Machine Learning Approach

#### 3.2.1. Estimation of UPDRS on the Basis of Eye Movement Measurements

#### 3.2.2. Estimation of UPDRS on the Basis of Eye Movement and Additional Measurements

- PDQ39—quality of life measure
- AIMS—Abnormal Involuntary Movement Score
- Epworth—sleep scale

#### 3.2.3. Estimation of Different Therapies Effects

**the global accuracy was 0.53**, coverage for decision classes: 0.5, 0.5, 0.75, 0.7.

**the global accuracy was 0.91**; coverage for decision classes: 0.6, 0.6, 0.6, 0.25.

**94.4%**level. The confusion matrix for this test can be found in Table 10.

#### 3.2.4. Estimation of UPDRS II and UPDRS III on the Basis of Eye Movement and Additional Measurements

#### 3.3. Comparison of Rough Set Based and Other Classifiers

## 4. Discussion

## 5. Conclusions

## Acknowledgments

## Author Contributions

## Conflicts of Interest

## Appendix A

Pat# | YearOfBirth | UPDRS_II | UPDRS_III | UPDRS_IV | UPDRS_Total | HYscale | SchwabEnglandScale | SccLat | SccDur | SccAmp | SccVel | PDQ39 | AIMS | Epworth | Session |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|

11 | 1955 | 25 | 45 | 3 | 75 | 4 | 60 | 259.1 | 51.9 | 14.5 | 552.2 | 43 | 2 | 8 | 1 |

11 | 1955 | 10 | 17 | 3 | 32 | 1 | 90 | 204.4 | 51.6 | 14.3 | 549.7 | 43 | 2 | 8 | 2 |

11 | 1955 | 25 | 32 | 3 | 62 | 4 | 60 | 314.6 | 46.2 | 12.2 | 514 | 43 | 2 | 8 | 3 |

11 | 1955 | 10 | 10 | 3 | 25 | 1 | 90 | 238.8 | 54.4 | 10.8 | 371.7 | 43 | 2 | 8 | 4 |

14 | 1979 | 17 | 43 | 0 | 61 | 3 | 60 | 173 | 52 | 9.19 | 393 | 31 | 0 | 8 | 1 |

14 | 1979 | 8 | 14 | 0 | 23 | 1.5 | 90 | 184 | 51 | 10.9 | 464 | 31 | 0 | 8 | 2 |

14 | 1979 | 17 | 32 | 0 | 50 | 3 | 60 | 331 | 45 | 9.1 | 506 | 31 | 0 | 8 | 3 |

14 | 1979 | 8 | 9 | 0 | 18 | 1.5 | 90 | 172 | 48 | 10.5 | 462 | 31 | 0 | 8 | 4 |

25 | 1956 | 20 | 40 | 2 | 62 | 3 | 70 | 301.8 | 49.1 | 10 | 386.1 | 29 | 0 | 10 | 1 |

25 | 1956 | 3 | 2 | 2 | 7 | 1 | 100 | 29 | 0 | 10 | 2 | ||||

25 | 1956 | 20 | 40 | 2 | 62 | 3 | 70 | 29 | 0 | 10 | 3 | ||||

25 | 1956 | 3 | 2 | 2 | 7 | 1 | 100 | 243.3 | 47.2 | 14.7 | 607.6 | 29 | 0 | 10 | 4 |

27 | 1954 | 25 | 63 | 4 | 94 | 4 | 60 | 474.7 | 43.3 | 10.1 | 475 | 78 | 4 | 12 | 1 |

27 | 1954 | 13 | 37 | 4 | 56 | 1.5 | 90 | 260.5 | 44.9 | 10.7 | 485 | 78 | 4 | 12 | 2 |

27 | 1954 | 25 | 53 | 4 | 84 | 4 | 60 | 411.2 | 32.7 | 8.5 | 559.6 | 78 | 4 | 12 | 3 |

27 | 1954 | 13 | 24 | 4 | 43 | 1.5 | 90 | 217.8 | 45.9 | 9.5 | 446.4 | 78 | 4 | 12 | 4 |

28 | 1959 | 17 | 40 | 0 | 58 | 2 | 60 | 401.7 | 43.2 | 12.2 | 566.9 | 68 | 0 | 12 | 1 |

28 | 1959 | 18 | 21 | 0 | 40 | 1 | 90 | 296.6 | 45.8 | 11.3 | 474.5 | 68 | 0 | 12 | 2 |

28 | 1959 | 11 | 27 | 7 | 46 | 1.5 | 80 | 227.1 | 48.6 | 10.4 | 431.2 | 68 | 0 | 12 | 3 |

28 | 1959 | 11 | 4 | 0 | 16 | 1 | 100 | 198.3 | 46.8 | 9 | 376.2 | 68 | 0 | 12 | 4 |

38 | 1957 | 19 | 28 | 2 | 49 | 2.5 | 70 | 285.1 | 42.2 | 14.2 | 675.2 | 63 | 3 | 5 | 1 |

38 | 1957 | 5 | 15 | 2 | 22 | 1.5 | 90 | 216.7 | 47.6 | 11.6 | 509.7 | 63 | 3 | 5 | 2 |

38 | 1957 | 19 | 16 | 2 | 37 | 2.5 | 70 | 380.4 | 42.8 | 14.4 | 638.9 | 63 | 3 | 5 | 3 |

38 | 1957 | 5 | 5 | 2 | 12 | 1.5 | 90 | 187.3 | 44.5 | 10 | 482.6 | 63 | 3 | 5 | 4 |

41 | 1980 | 18 | 37 | 0 | 55 | 2 | 80 | 267.7 | 41.9 | 8.8 | 383.7 | 23 | 0 | 7 | 1 |

41 | 1980 | 1 | 7 | 0 | 8 | 1 | 100 | 183.5 | 47.3 | 8.4 | 311.7 | 23 | 0 | 7 | 2 |

41 | 1980 | 18 | 35 | 0 | 53 | 2 | 80 | 244.4 | 43.3 | 14.4 | 656.7 | 23 | 0 | 7 | 3 |

41 | 1980 | 1 | 5 | 0 | 6 | 1 | 100 | 182 | 51 | 10.6 | 377.2 | 23 | 0 | 7 | 4 |

45 | 1944 | 15 | 42 | 0 | 62 | 3 | 60 | 255 | 44 | 8.8 | 396 | 63 | 0 | 7 | 1 |

45 | 1944 | 13 | 21 | 0 | 39 | 1.5 | 90 | 336 | 44 | 10.3 | 459 | 63 | 0 | 7 | 2 |

45 | 1944 | 15 | 27 | 0 | 47 | 3 | 60 | 331 | 43 | 9.3 | 406 | 63 | 0 | 7 | 3 |

45 | 1944 | 13 | 12 | 0 | 30 | 1.5 | 90 | 241 | 46 | 10.7 | 458 | 63 | 0 | 7 | 4 |

63 | 1960 | 23 | 26 | 2 | 54 | 3 | 70 | 227.6 | 42.3 | 15.8 | 701.5 | 89 | 0 | 12 | 1 |

63 | 1960 | 13 | 18 | 2 | 36 | 1.5 | 90 | 207.4 | 41.8 | 13.4 | 626.79 | 89 | 0 | 12 | 2 |

63 | 1960 | 23 | 26 | 2 | 54 | 3 | 70 | 187.7 | 44.3 | 14.2 | 604.79 | 89 | 0 | 12 | 3 |

63 | 1960 | 13 | 16 | 2 | 34 | 1.5 | 90 | 258.89 | 42.9 | 8.19 | 357.7 | 89 | 0 | 12 | 4 |

64 | 1957 | 24 | 49 | 2 | 77 | 4 | 60 | 230.3 | 46.6 | 7.7 | 382.7 | 104 | 15 | 1 | |

64 | 1957 | 29 | 2 | 33 | 2 | 80 | 104 | 15 | 2 | ||||||

64 | 1957 | 39 | 2 | 43 | 4 | 60 | 104 | 15 | 3 | ||||||

64 | 1957 | 12 | 23 | 2 | 39 | 2 | 80 | 194.1 | 43.2 | 7.2 | 326.89 | 104 | 15 | 4 |

## Appendix B

**Table B1.**Comparison of different classifiers with accuracy (ACC) and Matthews correlation coefficient (MCC) for different classifiers.

Fold# | Alglorithm | ACC | MCC |
---|---|---|---|

4 fold | Bayes | 0.425 | 0.25730532 |

4 fold | Decision Tree | 0.45 | 0.26164055 |

4 fold | Random Forest | 0.625 | 0.50898705 |

4 fold | Tree Ensemble | 0.65 | 0.53175848 |

4 fold | WEKA-Decision Table | 0.775 | 0.70116894 |

4 fold | WEKA-Random Forest | 0.45 | 0.27670875 |

40 fold | Bayes | 0.525 | 0.38539238 |

40 fold | Decision Tree | 0.62162 | 0.49525773 |

40 fold | Random Forest | 0.675 | 0.56408389 |

40 fold | Tree Ensemble | 0.725 | 0.63448742 |

40 fold | WEKA-Decision Table | 0.775 | 0.7007412 |

40 fold | WEKA-Random Forest | 0.55 | 0.38935439 |

5 fold | Bayes | 0.525 | 0.36809666 |

5 fold | Decision Tree | 0.41026 | 0.21908988 |

5 fold | Random Forest | 0.6 | 0.46007798 |

5 fold | Tree Ensemble | 0.7 | 0.60544001 |

5 fold | WEKA-Decision Table | 0.775 | 0.69943346 |

5 fold | WEKA-Random Forest | 0.6 | 0.46723502 |

6 fold | Bayes | 0.53 | 0.33183479 |

6 fold | Decision Tree | 0.51 | 0.17393888 |

6 fold | Random Forest | 0.68 | 0.55837477 |

6 fold | Tree Ensemble | 0.65 | 0.49267634 |

6 fold | WEKA-Decision Table | 0.48 | 0.7007412 |

6 fold | WEKA-Random Forest | 0.63 | 0.49387519 |

6 fold/SMOTE | Bayes | 0.95 | 0.9339666 |

6 fold/SMOTE | Decision Tree | 0.93082 | 0.90779989 |

6 fold/SMOTE | Random Forest | 0.975 | 0.96619735 |

6 fold/SMOTE | Tree Ensemble | 0.975 | 0.96675826 |

6 fold/SMOTE | WEKA-Decision Table | 0.9125 | 0.88469729 |

6 fold/SMOTE | WEKA-Random Forest | 0.95625 | 0.94205299 |

## Appendix C

**Table C1.**Comparison rough set classifiers with different fold number, accuracy (ACC), Matthews correlation coefficient (MCC).

Fold# | Dataset | Decision Class | Algorithm | ACC | MCC |
---|---|---|---|---|---|

6 fold | no RS | session | decomposition tree | 0.53 | 0.347106 |

5 fold | no RS | session | decomposition tree | 0.257 | 0.138529 |

4 fold | no RS | session | decomposition tree | 0.156 | 0.036866 |

40 fold | no RS | session | decision rules | 0.25 | 0.017689 |

6 fold | RS + PDQ | session | decomposition tree | 0.944 | 0.951147 |

5 fold | RS + PDQ | session | decomposition tree | 0.777 | 0.745525 |

4 fold | RS + PDQ | session | decomposition tree | 0.783 | 0.789729 |

40 fold | RS + PDQ | session | decision rules | 0.5 | 0.323653 |

6 fold | RS + PDQ | total UPDRS | decomposition tree | 0.903 | 0.854696 |

5 fold | RS + PDQ | total UPDRS | decomposition tree | 0.627 | 0.434957 |

4 fold | RS + PDQ | total UPDRS | decomposition tree | 0.818 | 0.780431 |

40 fold | RS + PDQ | total UPDRS | decision rules | 0.675 | 0.549481 |

6 fold | RS + PDQ | UPDRS II | decomposition tree | 0.62 | 0.244933 |

5 fold | RS + PDQ | UPDRS II | decomposition tree | 0.83 | 0.736768 |

4 fold | RS + PDQ | UPDRS II | decomposition tree | 0.83 | 0.699704 |

40 fold | RS + PDQ | UPDRS II | decision rules | 0.7 | 0.661798 |

6 fold | RS + PDQ | UPDRS III | decomposition tree | 0.817 | 0.583869 |

5 fold | RS + PDQ | UPDRS III | decomposition tree | 0.543 | 0.432082 |

4 fold | RS + PDQ | UPDRS III | decomposition tree | 0.506 | 0.33002 |

40 fold | RS + PDQ | UPDRS III | decision rules | 0.65 | 0.507846 |

6 fold | with RS | session | decomposition tree | 0.91 | 0.84503 |

5 fold | with RS | session | decomposition tree | 0.85 | 0.83524 |

4 fold | with RS | session | decomposition tree | 0.838 | 0.724788 |

40 fold | with RS | session | decision rules | 0.5 | 0.323653 |

## References

- Przybyszewski, A.W.; Kon, M.; Szlufik, S.; Dutkiewicz, J.; Habela, P.; Koziorowski, D.M. Data Mining and Machine Learning on the Basis from Reflexive Eye Movements Can Predict Symptom Development in Individual Parkinson’s Patients. In Nature-Inspired Computation and Machine Learning; Gelbukh, A., Espinoza, F.C., Galicia-Haro, S.N., Eds.; Springer International Publishing: Heidelberg, Germany, 2014; pp. 499–509. [Google Scholar]
- Przybyszewski, A.W.; Gaska, J.P.; Foote, W.; Pollen, D.A. Striate cortex increases contrast gain of macaque LGN neurons. Vis. Neurosci.
**2000**, 17, 485–494. [Google Scholar] [CrossRef] [PubMed] - Przybyszewski, A.W. The Neurophysiological Bases of Cognitive Computation Using Rough Set Theory. In Transactions on Rough Sets IX; Peters, J.F., Skowron, A., Rybiński, H., Eds.; Springer-Verlag: Berlin, Germany; Heidelberg, Germany, 2008; pp. 287–317. [Google Scholar]
- Przybyszewski, A.W. Logical Rules of Visual Brain: From Anatomy through Neurophysiology to Cognition. Cogn. Syst. Res.
**2010**, 11, 53–66. [Google Scholar] [CrossRef] - Pizzolato, G.; Mandat, T. Deep Brain Stimulation for Movement Disorders. Front. Integr. Neurosci.
**2012**, 6, 2–5. [Google Scholar] [CrossRef] [PubMed] - Pawlak, Z. Rough Sets; Springer Netherlands: Dordrecht, The Netherlands, 1991. [Google Scholar]
- Bazan, J.G.; Nguyen, H.S.; Nguyen, T.T.; Skowron, A.; Stepaniuk, J. Synthesis of Decision Rules for Object Classification. In Incomplete Information: Rough Set Analysis; Orłowska, P.D.E., Ed.; Physica-Verlag HD: Heidelberg, Germany, 1998; pp. 23–57. [Google Scholar]
- Bazan, J.G.; Szczuka, M. RSES and RSESlib—A Collection of Tools for Rough Set Computations. In Rough Sets and Current Trends in Computing; Ziarko, W., Yao, Y., Eds.; Springer: Heidelberg, Germany, 2001; pp. 106–113. [Google Scholar]
- Breiman, L. Random Forests. Mach. Learn.
**2001**, 45, 5–32. [Google Scholar] [CrossRef] - Kohavi, R. The power of decision tables. In Machine Learning: ECML-95; Lavrac, N., Wrobel, S., Eds.; Springer: Heidelberg, Germany, 1995; pp. 174–189. [Google Scholar]
- John, G.; Langley, P. Estimating Continuous Distributions in Bayesian Classifiers. In Proceedings of the Eleventh Conference on Uncertainty in Artificial Intelligence, San Francisco, CA, USA, 1995; pp. 338–345.
- Bazan, J.G.; Nguyen, H.S.; Nguyen, S.H.; Synak, P.; Wróblewski, J. Rough Set Algorithms in Classification Problem. In Rough Set Methods and Applications; Polkowski, P.L., Tsumoto, P.S., Lin, P.T.Y., Eds.; Physica-Verlag HD: Heidelberg, Germany, 2000; pp. 49–88. [Google Scholar]

**Figure 1.**An example of experimental recordings from Pat #38 in two sessions: upper plots from session S1: MedOFF & StimOFF; lower plots from session S4: MedON & StimON; left plots show latency measurements, right plots—saccade amplitude and velocity. Notice change in variability of responses between S1 and S4. These plots were obtained directly from the saccadometer (Ober Consulting), different colours are related to different eyes.

**Figure 2.**This graph shows parallel changes in UPDRS and reflexive saccade latencies as functions of medication and stimulation. Changes between control and MedOnStimOn were significantly different for UPDRS p < 0.001, RS p < 0.01.

P# | Age | Sex | t_dur | S# | UPDRS | HYsc | SccDur | SccLat | SccAmp | SccVel |
---|---|---|---|---|---|---|---|---|---|---|

28 | 54 | 1 | 8 | 1 | 58 | 2.0 | 43 | 402 | 12 | 566.9 |

28 | 54 | 1 | 8 | 2 | 40 | 1.0 | 46 | 297 | 11 | 474.5 |

28 | 54 | 1 | 8 | 2 | 40 | 1.0 | 49 | 227 | 10 | 431.2 |

28 | 54 | 1 | 8 | 4 | 16 | 1.0 | 47 | 198 | 9 | 376.2 |

38 | 56 | 0 | 11 | 1 | 49 | 2.5 | 42 | 285 | 14 | 675.2 |

38 | 56 | 0 | 11 | 2 | 22 | 1.5 | 48 | 217 | 12 | 509.7 |

38 | 56 | 0 | 11 | 3 | 37 | 2.5 | 43 | 380 | 14 | 638.9 |

38 | 56 | 0 | 11 | 4 | 12 | 1.5 | 45 | 187 | 10 | 482.6 |

P# | Age | t_dur | S# | HYsc | SccDur | SccLat | SccAmp | UPDRS |
---|---|---|---|---|---|---|---|---|

28 | (−Inf, 55.0) | * | 1 | * | (−Inf, 45.5) | (260.0, Inf) | (10.5, Inf) | (55.0, Inf) |

28 | (−Inf, 55.0) | * | 2 | * | (45.5, Inf) | (260.0, Inf) | (10.5, Inf) | (22.5, 55.0) |

28 | (−Inf, 55.0) | * | 2 | * | (45.5, Inf) | (−Inf, 260.0) | (−Inf, 10.5) | (22.5, 55.0) |

28 | (−Inf, 55.0) | * | 4 | * | (45.5, Inf) | (−Inf, 260.0) | (−Inf, 10.5) | (14.0, 22.5) |

38 | (55.0, Inf) | * | 1 | * | (−Inf, 45.5) | (260.0, Inf) | (10.5, Inf) | (22.5, 55.0) |

38 | (55.0, Inf) | * | 2 | * | (45.5, Inf) | (−Inf, 260.0) | (10.5, Inf) | (14.0, 22.5) |

38 | (55.0, Inf) | * | 3 | * | (−Inf, 45.5) | (260.0, Inf) | (10.5, Inf) | (22.5, 55.0) |

38 | (55.0, Inf) | * | 4 | * | (−Inf, 45.5) | (−Inf, 260.0) | (−Inf, 10.5) | (−Inf, 14.0) |

**Table 3.**Confusion matrix for UPDRS Total binned into four classes. Column headers represent predicted values and row headers represent actual values.

55.0, Inf | 22.5, 55.0 | −Inf, 14.0 | 14.0, 22.51 | ACC | |
---|---|---|---|---|---|

55.0, Inf | 0.3 | 0.3 | 0.0 | 0.0 | 0.2 |

22.5, 55.0 | 0.0 | 1.5 | 0.0 | 0.0 | 1.0 |

−Inf, 14.0 | 0.0 | 0.3 | 0.0 | 0.2 | 0.0 |

14.0, 22.5 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 |

TPR | 0.2 | 0.8 | 0.0 | 0.0 |

**Table 4.**Confusion matrix for total UPDRS prediction with PDQ, AIMS, and Epworth attributes. Column headers represent predicted values, and row headers represent actual values.

“(55, 94]” | “(25, 40]” | “[6, 25]” | “(40, 55]” | ACC | |
---|---|---|---|---|---|

“(55, 94]” | 0.5 | 0.17 | 0 | 0 | 0.417 |

“(25, 40]” | 0 | 0.5 | 0 | 0 | 0.5 |

“[6, 25]” | 0 | 0 | 1.17 | 0 | 1 |

“(40, 55]” | 0 | 0 | 0.17 | 0.33 | 0.333 |

TPR | 0.5 | 0.42 | 0.94 | 0.33 |

Pat# | Age | t_dur | SEngs | UPDRS III | UPDRS IV | UPDRS | Sess# |
---|---|---|---|---|---|---|---|

(27.5, 41.5) | (43.5, Inf) | * | (−Inf, 75) | (36.0, 46.0) | (10.5, Inf) | (1.75, Inf) | 1 |

(27.5, 41.5) | (43.5, Inf) | * | (75, Inf) | (13.0, 26.0) | (10.5, Inf) | (−Inf, 1.75) | 2 |

(27.5, 41.5) | (43.5, Inf) | * | (75, Inf) | (−Inf, 6.0) | (10.5, Inf) | (−Inf, 1.75) | 4 |

(27.5, 41.5) | (43.5, Inf) | * | (−Inf, 75) | (26.0, 36.0) | (10.5, Inf) | (1.75, Inf) | 1 |

(27.5, 41.5) | (43.5, Inf) | * | (75, Inf) | (13.0, 26.0) | (10.5, Inf) | (−Inf, 1.75) | 2 |

(27.5, 41.5) | (43.5, Inf) | * | (−Inf, 75) | (13.0, 26.0) | (10.5, Inf) | (1.75, Inf) | 3 |

(27.5, 41.5) | (43.5, Inf) | * | (75, Inf) | (−Inf, 6.0) | (10.5, Inf) | (−Inf, 1.75) | 4 |

**Table 6.**Confusion matrix for different session numbers (S1–S4). Column headers represent predicted values and row headers represent actual values.

1 | 2 | 3 | 4 | ACC | |
---|---|---|---|---|---|

1 | 0.5 | 0.0 | 0.5 | 0.0 | 0.3 |

2 | 0.0 | 0.5 | 0.0 | 0.3 | 0.4 |

3 | 0.8 | 0.0 | 0.2 | 0.0 | 0.2 |

4 | 0.0 | 0.5 | 0.0 | 0.5 | 0.4 |

TPR | 0.3 | 0.3 | 0.2 | 0.4 |

Pat# | Age | SccVel | UPDRS III | HYsc | SccDur | SccLat | SccAmp | Ses# |
---|---|---|---|---|---|---|---|---|

(27.5, 34.5) | * | (458.5, 578.0) | (36, Inf) | (1.75, Inf) | (38, Inf) | (308.5, Inf) | * | 1 |

(27.5, 34.5) | * | (458.5, 578) | (11.5, 36) | (−Inf, 1.75) | (38.0, Inf) | (−Inf, 308.5) | * | 2 |

(27.5, 34.5) | * | (341.5, 403) | (−Inf, 11.5) | (−Inf, 1.75) | (38, Inf) | (−Inf, 308.5) | * | 4 |

(34.5, Inf) | * | (665.5, Inf) | (11.5, 36.0) | (1.75, Inf) | (38.0, Inf) | (−Inf, 308.5) | * | 1 |

(34.5, Inf) | * | (458.5, 578) | (11.5, 36) | (−Inf, 1.75) | (38.0, Inf) | (−Inf, 308.5) | * | 2 |

(34.5, Inf) | * | (578.0, 665.5) | (11.5, 36) | (1.75, Inf) | (38.0, Inf) | (308.5, Inf) | * | 3 |

(34.5, Inf) | * | (458.5, 578) | (−Inf, 11.1) | (−Inf, 1.75) | (38, Inf) | (−Inf, 308.5) | * | 4 |

**Table 8.**Confusion matrix for different session numbers (S1–S4). Column headers represent predicted values and row headers represent actual values.

1 | 2 | 3 | 4 | ACC | |
---|---|---|---|---|---|

1 | 0.8 | 0.0 | 0.0 | 0.0 | 0.7 |

2 | 0.0 | 0.7 | 0.0 | 0.0 | 0.7 |

3 | 0.2 | 0.0 | 0.8 | 0.0 | 0.6 |

4 | 0.0 | 0.2 | 0.0 | 0.3 | 0.25 |

TPR | 0.7 | 0.7 | 0.6 | 0.25 |

**Table 9.**Part of the discretized decision table for session prediction. As we can see for this dataset only PDQ scale give significant value for prediction results.

Pat# | UPDRS_III | SEsc | SccLat | SccDur | SccAmp | PDQ39 | Ses# |
---|---|---|---|---|---|---|---|

25 | (36.0, 44.5) | (−Inf, 85.0) | (207.64, Inf) | (42.5, Inf) | (9.9, Inf) | (−Inf, 78.5) | 1 |

63 | (13.0, 26.0) | (85.0, Inf) | (207.64, Inf) | (42.5, Inf) | (−Inf, 9.9) | (78.5, Inf) | 4 |

25 | (36.0, 44.5) | (−Inf, 85.0) | MISSING | MISSING | MISSING | (−Inf, 78.5) | 3 |

25 | (−Inf, 6.0) | (85.0, Inf) | (207.64, Inf) | (42.5, Inf) | (9.9, Inf) | (−Inf, 78.5) | 4 |

63 | (26.0, 36.0) | (−Inf, 85.0) | (207.64, Inf) | (−Inf, 42.5) | (9.9, Inf) | (78.5, Inf) | 1 |

all the rest | (36.0, 44.5) | (85.0, Inf) | (207.64, Inf) | (42.5, Inf) | (9.9, Inf) | (−Inf, 78.5) | 2 |

all the rest | (44.5, Inf) | (−Inf, 85.0) | (207.64, Inf) | (−Inf, 42.5) | (−Inf, 9.9) | (−Inf, 78.5) | 3 |

**Table 10.**Table below shows the confusion matrix for extension of our previous results in prediction session. Global accuracy was established on 94.4% with 47% coverage. We used six-fold cross validation method with tree decomposition algorithm. Column headers represent predicted values, row headers represent actual values.

1 | 2 | 3 | 4 | ACC | |
---|---|---|---|---|---|

1 | 0.5 | 0 | 0 | 0 | 0.333 |

2 | 0 | 0.83 | 0 | 0 | 0.5 |

3 | 0 | 0 | 0.33 | 0 | 0.333 |

4 | 0 | 0.17 | 0 | 1 | 0.75 |

TPR | 0.33 | 0.42 | 0.33 | 0.83 |

Pat# | UPDRS II | Ses# | SccLat | PDQ39 | UPDRS III |
---|---|---|---|---|---|

{11, 14, 25, 27, 28, 41, 45, 64} | (14.5, Inf) | {‘S2’} | (259.695, Inf) | (65.5, Inf) | (14, 24] |

{11, 14, 25, 27, 28, 41, 45, 64} | (−Inf, 14.5) | {‘S3’} | (−Inf, 259.695) | (65.5, Inf) | (24, 35] |

{11, 14, 25, 27, 28, 41, 45, 64} | (−Inf, 14.5) | * | (−Inf, 259.695) | (65.5, Inf) | [2, 14] |

38 | (14.5, Inf) | {‘S1’} | (259.695, Inf) | (−Inf, 65.5) | (24, 35] |

38 | (−Inf, 14.5) | {‘S2’} | (−Inf, 259.695) | (−Inf, 65.5) | (14, 24] |

38 | (14.5, Inf) | {‘S3’} | (259.695, Inf) | (−Inf, 65.5) | (14, 24] |

63 | (14.5, Inf) | {‘S1’} | (−Inf, 259.695) | (65.5, Inf) | (24, 35] |

**Table 12.**Confusion matrix for UPDRS III model. Classification accuracy 81.7%, 61.1% coverage, 40 examination records. Test method: six-fold cross validation.

“(35, 63]” | “(14, 24]” | “(24, 35]” | “[2, 14]” | ACC | |
---|---|---|---|---|---|

“(35, 63]” | 1.33 | 0 | 0.33 | 0 | 0.667 |

“(14, 24]” | 0 | 0 | 0 | 0.17 | 0 |

“(24, 35]” | 0 | 0 | 0.67 | 0 | 0.667 |

“[2, 14]” | 0 | 0.33 | 0 | 0.83 | 0.417 |

TPR | 0.83 | 0 | 0.58 | 0.5 |

**Table 13.**Confusion matrix for UPDRS II classifier. Accuracy 78.6%, coverage 75.7%. 40 examination records classified with decomposition tree and six-fold cross validation method.

“(19, 25]” | “[1, 10]” | “(13, 19]” | “(10, 13]” | ACC | |
---|---|---|---|---|---|

“(19, 25]” | 1.33 | 0 | 0 | 0 | 0.833 |

“[1, 10]” | 0 | 0.83 | 0 | 0.33 | 0.5 |

“(13, 19]” | 0 | 0 | 1 | 0 | 0.5 |

“(10, 13]” | 0 | 0.5 | 0.33 | 0.17 | 0.056 |

TPR | 0.83 | 0.58 | 0.5 | 0.08 |

**Table 14.**As we can for those data sets we were able to achieve the most accurate results for the Rough Set-based approach but in most cases there were algorithms that could give predictions on a similar level, with the exception of classifiers for the session. Classifiers were tested on dataset with PDQ attribute.

Classifier | UPDRS Total | UPDRS II | UPDRS III | Sess# |
---|---|---|---|---|

Bayes | 53.17% | 77.38% | 32.54% | 0.00% |

Decision Tree | 50.79% | 7.94% | 62.30% | 30.16% |

Random Forest | 67.86% | 79.76% | 60.32% | 25.00% |

Tree Ensemble | 65.48% | 70.24% | 54.76% | 25.00% |

WEKA-Random Forest | 47.62% | 73.02% | 49.60% | 12.30% |

WEKA-Decision Table | 62.54% | 43.25% | 77.38% | 35.71% |

Rough Set | 90.30% | 78.60% | 81.70% | 94.40% |

© 2016 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC-BY) license (http://creativecommons.org/licenses/by/4.0/).

## Share and Cite

**MDPI and ACS Style**

Przybyszewski, A.W.; Kon, M.; Szlufik, S.; Szymanski, A.; Habela, P.; Koziorowski, D.M.
Multimodal Learning and Intelligent Prediction of Symptom Development in Individual Parkinson’s Patients. *Sensors* **2016**, *16*, 1498.
https://doi.org/10.3390/s16091498

**AMA Style**

Przybyszewski AW, Kon M, Szlufik S, Szymanski A, Habela P, Koziorowski DM.
Multimodal Learning and Intelligent Prediction of Symptom Development in Individual Parkinson’s Patients. *Sensors*. 2016; 16(9):1498.
https://doi.org/10.3390/s16091498

**Chicago/Turabian Style**

Przybyszewski, Andrzej W., Mark Kon, Stanislaw Szlufik, Artur Szymanski, Piotr Habela, and Dariusz M. Koziorowski.
2016. "Multimodal Learning and Intelligent Prediction of Symptom Development in Individual Parkinson’s Patients" *Sensors* 16, no. 9: 1498.
https://doi.org/10.3390/s16091498