# Machine Learning of Infant Spontaneous Movements for the Early Prediction of Cerebral Palsy: A Multi-Site Cohort Study

^{1}

^{2}

^{3}

^{4}

^{5}

^{6}

^{7}

^{8}

^{9}

^{10}

^{11}

^{12}

^{13}

^{14}

^{*}

## Abstract

**:**

## 1. Introduction

## 2. Experimental Section

#### 2.1. Study Participants

#### 2.2. The Computer-Based Infant Movement Assessment (CIMA) Model

#### 2.2.1. Infant Motion Detection in Video Recording

#### 2.2.2. Movement Feature Extraction

#### 2.2.3. CP Prediction Model and Validity of the Model

#### 2.3. Observational GMA, Cerebral Imaging, Cerebral Palsy and Gross Motor Function

#### 2.4. Statistics of the Outcome of the CIMA Model

_{SD}) [12,30]. Kruskal–Wallis’s test with the post hoc Wilcoxon rank-sum test including Bonferroni correction assessed the significance of the difference in the proportion of CP risk-related movements between the different FM categories assessed by the observational GMA. The Wilcoxon rank-sum test was also used to assess the significance of the difference in the proportion of CP risk-related movements between infants developing CP with GMFCS I, II, or III (i.e., ambulatory CP) and those developing CP with GMFCS IV or V (i.e., non-ambulatory CP). All analyses and statistics were performed in Matlab 2018a and p-values below 0.05 were considered statistically significant.

## 3. Results

#### Proportion of Periods with CP Risk-Related Movements, CP Status and Gross Motor Function

_{SD}). The CIMA model had the best sensitivity, NPV and AUC. However, the specificity and PPV were slightly lower than for the GMA and neuroimaging results. The statistics in Table 3 are dependent on a decision threshold of 50%. The ROC curve and alternative thresholds are provided in Appendix E and cross-tables and mean square contingency coefficients are provided in Appendix F.

## 4. Discussion

_{SD}) used in several previous computer-based studies by our group [12,30]. The previously developed C

_{SD}was based on a frame differencing method which may be susceptible to differences in contrasts, light, and infant clothing, which may vary more in this larger multi-site cohort of infants. Furthermore, as the sample size and heterogeneity of children with CP increase, it becomes more challenging for a single predefined feature, such as C

_{SD}, to contain information of various characteristics of the infant movement repertoire relevant for a clinical outcome such as CP. Thus, we argue that it is likely that the predictive performance of other suggested single features such as relative movement frequency [15] and mean and minimum velocity [30] will potentially decay in larger multi-site populations of high-risk infants. The performance of the presented CIMA model suggests that overall variables, such as the proportion (%) of periods with CP risk-related movements, should be based on a cluster of movement features rather than single “key” features.

## 5. Conclusions

## Author Contributions

## Funding

## Acknowledgments

## Conflicts of Interest

## Appendix A. Detailed Inclusion Criteria and Characteristics of the Multi-Site Cohort

Risk Group | N (%) |
---|---|

GA < 28 weeks and/or BW ≤ 1000 g | 167 (44.3) |

- Boys | 90 (53.9) |

- GA (weeks), mean (SD) | 26.3 (1.7) |

- BW (g), mean (SD) | 833 (178) |

GA 28–36 weeks and BW > 1001 g | 59 (15.6) |

Neonatal arterial ischemic stroke | 15 (4.0) |

Neonatal encephalopathy | 50 (13.3) |

CHD w/surgery before 4 weeks | 39 (10.3) |

Other ^{a} | 47 (12.5) |

^{a}: Infants who were referred to neurodevelopmental follow-up at discharge from the neonatal intensive care unit due to significant abnormalities on cerebral imaging (intraventricular hemorrhages III–IV, other intracranial hemorrhages with or without seizures, cystic periventricular leukomalacia, ventriculomegaly, venous sinus thrombosis), central nervous system infection, medically complex infants (syndromes/chromosomal abnormalities, multiple congenital anomalies, hydrops fetalis, severe lung hypoplasia, protracted hypoglycemia, seizures with unknown etiology) and severe intrauterine growth restriction. One second twin came to follow-up due to referral of the first twin.

## Appendix B. Feature Extraction of Complex and Variable Movements

**r**(t) or

**d**(t) for iterative steps) along the unit direction vectors θ

_{k}of the m-dimensional sphere.

**m**(t) of all envelope curves, ${e}_{{\theta}_{k}}\left(t\right)$, across all m directions of the sphere by the following equation:

_{1}(t) around the mean m

_{1}(t) is defined as d

_{1}(t) = x(t) − m

_{1}(t). If d

_{1}(t) satisfies the selected stopping criteria, then d

_{1}(t) is defined as an intrinsic movement modality of the infant (i.e., intrinsic mode function; IMF) and Steps 2 to 5 is performed on first residual, r

_{1}(t) = x(t) − d

_{1}(t). The second IMF is defined as d

_{2}(t) = r

_{1}(t) − m

_{2}(t) with residual r

_{2}(t) = r

_{1}(t) − d

_{2}(t). Consequently, the nth IMF is defined as d

_{n}(t) = r

_{n}

_{−1}(t) − m

_{n}(t) with residual r

_{n}(t) = r

_{n}

_{−1}(t) − d

_{n}(t). This iterative shifting procedure (i.e., Steps 2 to 5) is continued until two maxima ${p}_{{\theta}_{k}}^{max}\left(t\right)$ of the projection ${p}_{{\theta}_{k}}\left(t\right)$ in Step 3 can no longer be found. If d

_{n}(t) does not satisfy the stopping criteria, Steps 2 to 5 are performed as an iterative procedure on d

_{n}(t) until the stopping criteria is met and an IMF is defined. Subsequently, Steps 2 to 5 are repeated on the residual, r

_{n}(t) = r

_{n}

_{−1}(t) − d

_{n}(t). The stopping criteria used in the present study is similar to the stopping criteria proposed by Rilling et al. [34], except that we excluded the criteria of equality between the number of zero crossings and number of maxima. The sum of all IMFs and the final residual, $\mathrm{x}\left(t\right)={\displaystyle \sum}_{n=1}^{N}{\mathrm{d}}_{n}\left(t\right)+{\mathrm{r}}_{N}\left(t\right)$ correspond to the infant segment movements x(t), where N is the number of IMFs. In the present study, N = 11 for all video recordings.

_{n}(t) = [a

_{n,1}(t), a

_{n,2}(t),…,a

_{n,12}(t)] and f

_{n}(t) = [f

_{n,1}(t), f

_{n,2}(t),…,f

_{n,12}(t)] are the vector of instantaneous amplitude and frequency of d

_{n}(t) = [d

_{n,1}(t), d

_{n,2}(t),…,d

_{n,12}(t)] where each element is the instantaneous amplitude a

_{i}(t) and frequency f

_{i}(t) of d

_{n,i}(t) of a single segment in horizontal or vertical direction. The spectral density S

_{n,ii}of a

_{n,i}(t) was estimated as:

_{n,i}(t) was estimated as:

_{n,ii}(t) and f

_{n,i}(t) were divided into 5 second non-overlapping time windows and the sum of S

_{n,i}(t) (i.e., total spectral energy) and mean f

_{n}(t) was computed for each window. This provided 264 features (i.e., number of IMFs x number of movement coordinates) for each time window. In addition, the instantaneous covariance S

_{n,ij}(t) for movement coordinates for body segment/direction i and j of each of the N scales was estimated as:

_{n,ij}(t) was divided into 5 second non-overlapping time windows. The sum of S

_{n,ij}(t) (i.e., total spectral covariance) was assessed for each time window, resulting in 736 covariance combinations per time window. In total, 990 features were defined for each time window.

## Appendix C. CP Prediction Model

**X**be the scaled and centered feature matrix with 990 columns obtained by the procedure of Appendix B. Let

**Y**be a column vector with −1/1 elements according to CP outcome (i.e., CP = 1 and non-CP = −1) for each 5 second time window where the length of

**Y**is equal to the number of rows in X. The partial least square regression of

**X**and

**Y**was computed by the following nonlinear iterative partial least squares (NIPALS) algorithm [22]:

**w**are defined by:

**u**=

**Y**, for the first iteration of Step 1 to 5.

**c**are defined by

**t**of Equation A8. The first component is removed from

**X**and

**Y**by the following two equations:

**X**according to an inner 5-fold cross-validation procedure (see Section 2.2.3 in the main text). The obtained matrix

**T**contains N columns for containing X-scores

**t**in Equation A8 for each of the N repetitions of Steps 1 to 7.

**T**defines the CP risk-related components of the original feature matrix

**X**relevant for predicting CP outcome. Next, matrix

**T**is an input in a linear discriminative analysis (LDA) to obtain a single score ${Y}_{est}$ in the range [−1, 1] for each 5 second window where ${Y}_{est}>0$ indicated a CP risk-related movement in the 5 second window. The following Bayesian approximation of LDA was used to define ${Y}_{est}$:

**A’**is all elements except the last element a in $A={\left({Z}^{T}Z\right)}^{-1}{Z}^{T}B$, where $B=1$. Matrix

**Z**is equal to $\left[T,Y\right]$ for elements where

**Y**> 0 and $\left[-T,Y\right]$ for elements where $Y\le 0$ (i.e., last column of

**Z**equal to

**Y**).

## Appendix D. Cross-Validation Procedure

**Figure A1.**Schematic illustration of the cross-validation (CV) procedure used to validate the CIMA model.

## Appendix E. ROC Curve and Decision Thresholds

**Table A2.**Performance of CP prediction with different decision thresholds (%) for the proportion of periods with CP risk-related movements. The sensitivity, specificity, positive and negative predictive values and area under the curve (AUC) with 95% confidence intervals in brackets for the prediction of CP.

Threshold (%) | Sens. (%) | Spec. (%) | PPV (%) | NPV (%) |
---|---|---|---|---|

50 | 92.7 [80.1, 98.5] | 81.6 [77.0, 85.5] | 38.0 [32.5, 43.8] | 98.9 [96.8, 99.6] |

55 | 92.7 [80.1, 98.5] | 83.9 [79.6, 87.7] | 41.3 [35.2, 47.7] | 99.0 [96.9, 99.6] |

60 | 85.4 [70.8, 94.4] | 86.6 [82.5, 90.1] | 43.6 [36.6, 51.2] | 98.0 [95.9, 99.0] |

65 | 78.1 [62.4, 89.4] | 89.3 [85.5, 92.4] | 47.0 [38.5, 55.8] | 97.1 [94.9, 98.4] |

70 | 68.3 [51.9, 81.9] | 92.3 [88.9, 94.9] | 51.9 [41.3, 62.2] | 96.0 [93.8, 97.4] |

**Figure A2.**Receiver operating characteristic (ROC) curve of the proportion of periods with CP risk-related movements (red graph) compared to the ROC curve of the standard deviation of the center of motion (C

_{SD}) (blue graph). The horizontal and vertical dashed lines indicate points on the red ROC curve for the different decision thresholds represented in Table A2.

## Appendix F. Correlation between Imaging, GMA, and CIMA

Imaging/CIMA | CP Risk-Related Move > 50% | CP Risk-Related Move < 50% |
---|---|---|

Abnormal MR | 42 | 42 |

Normal MR | 57 | 234 |

GMA/CIMA | CP Risk-Related Move > 50% | CP Risk-Related Move < 50% |
---|---|---|

Abnormal FMs | 44 | 42 |

Normal FMs | 56 | 235 |

Imaging/GMA | Abnormal FMs | Normal FMs |
---|---|---|

Abnormal MR | 35 | 49 |

Normal MR | 50 | 241 |

## References

- Rosenbaum, P.; Paneth, N.; Leviton, A.; Goldstein, M.; Bax, M.; Damiano, D.; Dan, B.; Jacobsson, B. A report: The definition and classification of cerebral palsy. Dev. Med. Child Neurol. Suppl.
**2007**, 109, 8–14. [Google Scholar] - Oskoui, M.; Coutinho, F.; Dykeman, J.; Jette, N.; Pringsheim, T. An update on the prevalence of cerebral palsy: A systematic review and meta-analysis. Dev. Med. Child Neurol.
**2013**, 55, 509–519. [Google Scholar] [CrossRef] [PubMed] - Novak, I.; Morgan, C.; Adde, L.; Blackman, J.; Boyd, R.N.; Brunstrom-Hernandez, J.; Cioni, G.; Damiano, D.; Darrah, J.; Eliasson, A.C.; et al. Early, accurate diagnosis and early intervention in cerebral palsy: Advances in Diagnosis and Treatment. JAMA Pediatr.
**2017**, 171, 897–907. [Google Scholar] [CrossRef] [PubMed] - Guttmann, K.; Flibotte, J.; DeMauro, S.B. Parental Perspectives on Diagnosis and Prognosis of Neonatal Intensive Care Unit Graduates with Cerebral Palsy. J. Pediatr.
**2018**, 203, 156–162. [Google Scholar] [CrossRef] - Baird, G.; McConachie, H.; Scrutton, D. Parents’ perceptions of disclosure of the diagnosis of cerebral palsy. Arch. Dis. Child
**2000**, 83, 475–480. [Google Scholar] [CrossRef] [PubMed][Green Version] - Herskind, A.; Greisen, G.; Nielsen, J. Early identification and intervention in cerebral palsy. Dev. Med. Child Neurol.
**2014**, 57, 29–36. [Google Scholar] [CrossRef] [PubMed] - Hadders-Algra, M.; Gramsbergen, A. Discussion on the clinical relevance of activity-dependent plasticity after an insult to the developing brain. Neurosci. Biobehav. Rev.
**2007**, 31, 1213–1219. [Google Scholar] [CrossRef] - Maitre, N. Skeptism, cerebral palsy, and the general movement assessment. Dev. Med. Child Neurol.
**2018**. [Google Scholar] [CrossRef][Green Version] - Marcroft, C.; Khan, A.; Embleton, N.D.; Trenell, M.; Plotz, T. Movement recognition technology as a method of assessing spontaneous general movements in high risk infants. Front. Neurol.
**2014**, 5, 284. [Google Scholar] [CrossRef] - Cabon, S.; Poree, F.; Simon, A.; Rosec, O.; Pladys, P.; Carrault, G. Video and audio processing in paediatrics: A review. Physiol. Meas.
**2019**, 40, 02TR. [Google Scholar] [CrossRef] - Marchi, V.; Hakala, A.; Knight, A.; D’Acunto, F.; Scattoni, M.L.; Guzzetta, A.; Vanhatalo, S. Automated pose estimation captures key aspects of General Movements at eight to 17 weeks from conventional videos. Acta Paediatr.
**2019**. [Google Scholar] [CrossRef] [PubMed][Green Version] - Adde, L.; Helbostad, J.L.; Jensenius, A.R.; Taraldsen, G.; Grunewaldt, K.H.; Støen, R. Early prediction of cerebral palsy by computer-based video analysis of general movements: A feasibility study. Dev. Med. Child Neurol.
**2010**, 52, 773–778. [Google Scholar] [CrossRef] [PubMed] - Rahmati, H.; Aamo, O.M.; Stavdahl, Ø.; Dragon, R.; Adde, L. Video-based early cerebral palsy prediction using motion segmentation. In Proceedings of the 2014 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, Chicago, IL, USA, 26–30 August 2014; pp. 3779–3783. [Google Scholar]
- Rahmati, H.; Martens, H.; Aamo, O.M.; Stavdahl, Ø.; Støen, R.; Adde, L. Frequency Analysis and Feature Reduction Method for Prediction of Cerebral Palsy in Young Infants. IEEE Trans. Neural. Syst. Rehabil. Eng.
**2016**, 24, 11. [Google Scholar] [CrossRef] [PubMed] - Stahl, A.; Schellewald, C.; Stavdahl, Ø.; Aamo, O.M.; Adde, L.; Kirkerod, H. An optical flow-based method to predict infantile cerebral palsy. IEEE Trans. Neural. Syst. Rehabil. Eng.
**2012**, 20, 605–614. [Google Scholar] [CrossRef] [PubMed] - Orlandi, S.; Raghuram, K.; Smith, C.R.; Mansueto, D.; Church, P.; Shah, V.; Luther, M.; Chau, T. Detection of Atypical and Typical Infant Movement using Computer-based Video Analysis. In Proceedings of the 2018 40th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Honolulu, HI, USA, 18–21 July 2018. [Google Scholar]
- Kanemaru, N.; Watanabe, H.; Kihara, H.; Nakano, H.; Nakamura, T.; Nakano, J.; Taga, G.; Konishi, Y. Jerky spontaneous movements at term age in preterm infants who later developed cerebral palsy. Early Hum. Dev.
**2014**, 90, 387–392. [Google Scholar] [CrossRef] - Hadders-Algra, M. Neural substrate and clinical significance of general movements: An update. Dev. Med. Child Neurol.
**2018**, 60, 39–46. [Google Scholar] [CrossRef][Green Version] - Støen, R.; Boswell, L.; de Regnier, R.-A.; Fjørtoft, T.; Gaebler-Spira, D.; Ihlen, E.; Labori, C.; Loennecken, M.; Msall, M.; Möinichen, U.I.; et al. The predictive accuracy of the General Movement Assessment for cerebral palsy: A prospective, observational study of high-risk infants in a clinical follow-up setting. J. Clin. Med.
**2019**, 8(11), 1790. [Google Scholar] [CrossRef][Green Version] - Rahmati, H.; Dragon, R.; Aamo, O.M.; Adde, L.; Stavdahl, Ø. Weakly supervised motion segmentation with particle matching. Comput. Vis. Image Underst.
**2015**, 140, 30–42. [Google Scholar] [CrossRef] - Rehman, N.; Mandic, D.P. Multivariate empirical mode decomposition. Proc. R. Soc. A
**2010**, 466, 1291–1302. [Google Scholar] [CrossRef] - Huang, N.E.; Shen, Z.; Long, S.R.; Wu, M.C.; Shih, H.H.; Zheng, Q.; Yen, N.C.; Tung, C.C.; Liu, H.H. The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis. Proc. R. Soc. Lond. Ser. A Math. Phys. Eng. Sci.
**1998**, 454, 903–993. [Google Scholar] [CrossRef] - Cans, C. Surveillance of cerebral palsy in Europe: A collaboration of cerebral palsy surveys and registers. Surveillance of Cerebral Palsy in Europe (SCPE). Dev. Med. Child Neurol.
**2000**, 42, 816–824. [Google Scholar] [CrossRef] - Wold, S.; Sjöström, L.; Erikson, L. PLS regression: A basic tool of chemometrics. Chemom. Intell. Lab. Syst.
**2001**, 58, 37–52. [Google Scholar] [CrossRef] - Tang, L.; Peng, S.; Bi, Y.; Shan, P.; Hu, X. A New Method Combining LDA and PLS for Dimension Reduction. PLoS ONE
**2014**, 9, e96944. [Google Scholar] [CrossRef] [PubMed][Green Version] - Einspieler, C.; Prechtl, H.F.; Bos, A.; Ferrari, F.; Cioni, G. Prechtl’s Method on the Qualitative Assessment of General Movements in Preterm, Term and Young Infants; Mac Keith Press: London, UK, 2004. [Google Scholar]
- Einspieler, C.; Peharz, R.; Marschik, P.B. Fidgety movements—Tiny in appearance, but huge in impact. J. Pediatr. (Rio J)
**2016**, 92, 64–70. [Google Scholar] [CrossRef] [PubMed][Green Version] - Palisano, R.J.; Rosenbaum, P.; Walter, S.; Russell, D.; Wood, E.; Galuppi, B. Development and reliability of a system to classify gross motor function in children with cerebral palsy. Dev. Med. Child Neurol.
**1997**, 39, 214–223. [Google Scholar] [CrossRef] [PubMed] - Palisano, R.J.; Hanna, S.E.; Rosenbaum, P.L.; Russell, D.J.; Walter, S.D.; Wood, E.P.; Raina, P.S.; Galuppi, B.E. Validation of a model of gross motor function for children with cerebral palsy. Phys. Ther.
**2000**, 80, 974–985. [Google Scholar] - Støen, R.; Songstad, N.T.; Silberg, I.E.; Fjørtoft, T.; Jensenius, A.R.; Adde, L. Computer-based video analysis identifies infants with absence of fidgety movements. Pediatr. Res.
**2017**, 82, 665–679. [Google Scholar] [CrossRef][Green Version] - Einspieler, C.; Bos, A.F.; Krieber-Tomantschger, M.; Alvarado, E.; Barbosa, V.M.; Bertoncelli, N.; Burger, M.; Chorna, O.; Del Secco, S.; DeRegnier, R.A.; et al. Cerebral Palsy: Early markers of clinical phenotypes and functional outcome. J. Clin. Med.
**2019**, 8, 1616. [Google Scholar] [CrossRef][Green Version] - Cao, Z.; Hidalgo, G.; Simon, T.; Wei, S.-E.; Sheikh, Y. OpenPose: Realtime multi-person 2D pose estimation using part affinity fields. arXiv
**2018**, arXiv:1812.08008. [Google Scholar] [CrossRef][Green Version] - Fjørtoft, T.; Einspieler, C.; Adde, L.; Strand, L.I. Inter-observer reliability of the Assessment of Motor Repertoire 3 to 5 Months based on video recordings of infants. Early Hum. Dev.
**2009**, 85, 297–302. [Google Scholar] [CrossRef][Green Version] - Rilling, G.; Flandrin, P.; Gonçalves, P. On empirical mode decomposition and its algorithms. In Proceedings of the IEEE-EURASIP, Workshop on Nonlinear Signal and Image Processing NSIP-03, Grado, Italy, 8–10 June 2003. [Google Scholar]
- Marple, S.L. Computing the Discrete-Time Analytic Signal via FFT. IEEE Trans. Signal Process.
**1999**, 47, 2600–2603. [Google Scholar] [CrossRef]

**Figure 1.**Flow-chart of exclusion of video recordings for the development and testing of the Computer-based Infant Movement Assessment (CIMA) model.

**Figure 2.**Steps of the CIMA model. First, infant movements are detected by motion tracking of six body parts (head, trunk, arms, and legs) in the video. Second, features for the movement frequencies, amplitude, and covariation of the different body parts are extracted from the body part movement trajectories and used in the CP prediction model. The CP prediction model identifies 5 second periods with CP risk-related movements. Finally, the proportion (%) of periods with CP risk-related movements typically found in infants with CP is summarized and communicated as a CP risk indicator.

**Figure 3.**Each bar represents the proportion (%) of periods with CP risk-related movements represented in the video recordings of each of the 377 infants. The bars are centered around the decision threshold of 50% (horizontal line) for increased risk of CP. The red bars are from infants with confirmed CP diagnosis, whereas the blue bars represent the infants with a confirmed non-CP diagnosis.

**Figure 4.**Boxplot of the proportion of periods with CP risk-related movements assessed by the CIMA model (y-axis) and temporal organization of FMs assessed by observational GMA (x-axis) according to CP outcome. The red line indicates the median and blue box the interquartile range. The whiskers in dashed lines are 1.5 times the interquartile range and cover 99.3% of the data if normally distributed. Outliers are marked as red crosses. The horizontal dashed line represents a decision threshold of 50% for the CIMA model. FM− = absent FM; FM−/+ = sporadic FM; FM+ = intermittent FM; FM++ = continual FM; FMa = FM with exaggerated speed and amplitude.

**Table 1.**Summary of results in previous studies for the prediction of cerebral palsy (CP) with video-based automated infant movement analysis.

Study | Sample Size ^{1} | Sens. (%) | Spec. (%) | Acc (%) | Features |
---|---|---|---|---|---|

Adde [12] | 30 (13) | 85 | 88 | 88 * | C_{SD}, QoM |

Rahmati [13] | 78 (14) | 50 | 95 | 87 | FFT features |

Rahmati [14] | 78 (14) | 86 | 92 | 91 | FFT features |

Stahl [15] | 82 (15) | 85 | 95 | 94 | Wavelet features |

Orlandi [16] | 127 (16) | 44 | 99 | 92 | FFT/time features |

^{1}Sample size and number of infants with later CP diagnosis in parenthesis (..). * Value is area under receiver operating characteristic (ROC) curve. FFT = fast Fourier transformation (i.e., amplitude and frequency of infant movements); C

_{SD}= standard deviation of the center of motion; QoM = quantity-of-motion.

CP Status | N (%) |
---|---|

CP subtype * | |

Unilateral spastic | 8 (20) |

Bilateral spastic | 25 (61) |

Dyskinetic | 5 (12) |

Ataxic | 1 (2) |

Gross motor function (GMFCS) | |

GMFCS I | 11 (27) |

GMFCS II | 3 (7) |

GMFCS III | 6 (15) |

GMFCS IV | 10 (24) |

GMFCS V | 11 (27) |

**Table 3.**The sensitivity, specificity, positive and negative predictive values and area under the curve (AUC) with 95% confidence intervals in brackets for the prediction of CP.

Method | Sens. (%) | Spec. (%) | PPV (%) | NPV (%) | AUC * |
---|---|---|---|---|---|

CIMA | 92.7 [80.1, 98.5] | 81.6 [77.0, 85.5] | 38.0 [32.5, 43.8] | 98.9 [96.8, 99.6] | 0.87 [0.81, 0.91] |

GMA [19] | 76.2 [60.6, 88.0] | 82.4 [78.1, 86.2] | 33.3 [27.4, 39.8] | 96.8 [94.6, 98.1] | 0.82 [0.78, 0.85] |

Imaging [19] | 81.0 [65.9, 91.4] | 85.3 [81.2, 88.8] | 39.1 [32.5, 46.1] | 97.5 [95.4, 98.6] | 0.85 [0.81, 0.88] |

C_{SD} | 56.1 [39.8, 71.5] | 58.6 [53.2, 64.0] | 14.2 [10.9, 18.6] | 91.6 [88.5, 94.0] | 0.56 [0.48, 0.64] |

© 2019 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**

Ihlen, E.A.F.; Støen, R.; Boswell, L.; de Regnier, R.-A.; Fjørtoft, T.; Gaebler-Spira, D.; Labori, C.; Loennecken, M.C.; Msall, M.E.; Möinichen, U.I.;
et al. Machine Learning of Infant Spontaneous Movements for the Early Prediction of Cerebral Palsy: A Multi-Site Cohort Study. *J. Clin. Med.* **2020**, *9*, 5.
https://doi.org/10.3390/jcm9010005

**AMA Style**

Ihlen EAF, Støen R, Boswell L, de Regnier R-A, Fjørtoft T, Gaebler-Spira D, Labori C, Loennecken MC, Msall ME, Möinichen UI,
et al. Machine Learning of Infant Spontaneous Movements for the Early Prediction of Cerebral Palsy: A Multi-Site Cohort Study. *Journal of Clinical Medicine*. 2020; 9(1):5.
https://doi.org/10.3390/jcm9010005

**Chicago/Turabian Style**

Ihlen, Espen A. F., Ragnhild Støen, Lynn Boswell, Raye-Ann de Regnier, Toril Fjørtoft, Deborah Gaebler-Spira, Cathrine Labori, Marianne C. Loennecken, Michael E. Msall, Unn I. Möinichen,
and et al. 2020. "Machine Learning of Infant Spontaneous Movements for the Early Prediction of Cerebral Palsy: A Multi-Site Cohort Study" *Journal of Clinical Medicine* 9, no. 1: 5.
https://doi.org/10.3390/jcm9010005