📚 Machine Learning for Parkinson’s Disease Diagnosis | 机器学习与帕金森病诊断模型
Machine learning (ML) has emerged as a powerful tool in medical diagnosis, enabling computers to learn patterns from patient data and support clinical decision-making. Among its many applications, diagnosing Parkinson’s disease—a progressive neurological disorder affecting over 10 million people worldwide—stands out as a compelling case study. This article explores how ML models analyse voice recordings and other biomarkers to distinguish Parkinson’s patients from healthy individuals, and how these techniques fit within the A-Level Computer Science curriculum.
机器学习（ML）已成为医学诊断领域中的强大工具，它让计算机能够从患者数据中学习模式并辅助临床决策。在其众多应用场景中，诊断帕金森病——一种影响全球超过一千万人的进行性神经系统疾病——作为一个引人注目的案例脱颖而出。本文将探讨机器学习模型如何分析嗓音录音及其他生物标志物，以区分帕金森患者与健康个体，以及这些技术如何融入A-Level计算机科学课程体系。
1. Understanding Parkinson’s Disease | 认识帕金森病
Parkinson’s disease is a neurodegenerative condition characterised by the progressive loss of dopamine-producing neurons in the substantia nigra region of the brain. Common symptoms include tremors, rigidity, bradykinesia (slowness of movement), and postural instability. Notably, approximately 90% of Parkinson’s patients exhibit some form of vocal impairment, making voice analysis a promising non-invasive diagnostic pathway.
帕金森病是一种神经退行性疾病，其特征是大脑黑质区域中产生多巴胺的神经元逐渐丧失。常见症状包括震颤、肌肉强直、运动迟缓（动作缓慢）和姿势不稳。值得注意的是，大约90%的帕金森患者表现出某种形式的发声障碍，这使得嗓音分析成为一种前景广阔的无创诊断途径。
Traditional diagnosis relies on clinical observation and neurological examinations, which can be subjective and may miss early-stage cases. By the time motor symptoms become apparent, significant neural damage has already occurred. This motivates the use of computational methods that can detect subtle patterns in biological signals before overt symptoms emerge.
传统诊断依赖临床观察和神经学检查，这往往带有主观性，且可能遗漏早期病例。当运动症状变得明显时，显著的神经损伤已经发生。这促使我们使用计算方法在明显症状出现之前检测生物信号中的细微模式。
2. Machine Learning Fundamentals | 机器学习基础概念
Machine learning is a branch of artificial intelligence that enables systems to learn from data without being explicitly programmed. In supervised learning, the algorithm receives labelled training data—input-output pairs—and learns a mapping function from inputs to outputs. For Parkinson’s diagnosis, the input is a set of voice measurement features, and the output is a binary label: ‘patient’ (1) or ‘healthy’ (0).
机器学习是人工智能的一个分支，它使系统无需显式编程即可从数据中学习。在监督学习中，算法接收带标签的训练数据——即输入-输出对——并学习从输入到输出的映射函数。对于帕金森诊断而言，输入是一组嗓音测量特征，输出是二分类标签：“患者”（1）或“健康”（0）。
Classification problems such as this can be tackled using a variety of algorithms, including decision trees, support vector machines (SVM), k-nearest neighbours (k-NN), and artificial neural networks (ANN). Each algorithm has different strengths regarding interpretability, handling of non-linear relationships, and resistance to overfitting.
诸如此类的分类问题可以使用多种算法来解决，包括决策树、支持向量机（SVM）、k近邻（k-NN）和人工神经网络（ANN）。每种算法在可解释性、处理非线性关系的能力以及抗过拟合性能方面各有优势。
3. The Dataset: Voice Biomarkers | 数据集：嗓音生物标志物
A widely studied resource for Parkinson’s diagnosis is the UCI Parkinson’s Disease Classification dataset. It contains 195 sustained vowel phonations from 31 subjects, of whom 23 have Parkinson’s disease. Each recording is represented by 22 voice features plus a binary status label. These features fall into several categories: frequency measures, jitter, shimmer, and harmonic-to-noise ratios.
帕金森诊断研究中使用广泛的一个资源是UCI帕金森病分类数据集。该数据集包含31名受试者（其中23人患有帕金森病）的195次持续元音发声记录。每条录音由22个嗓音特征加一个二分类状态标签表示。这些特征分为几个类别：频率测量、抖动（jitter）、光泽度（shimmer）和谐波噪声比。
Frequency features include the average fundamental frequency (MDVP:Fo), the maximum (Fhi) and minimum (Flo) frequencies. Jitter measures capture cycle-to-cycle variations in pitch period, such as Jitter(%) and Jitter:RAP, while shimmer measures quantify amplitude instability across consecutive voice cycles. Additional features such as the noise-to-harmonics ratio (NHR) and recurrence period density entropy (RPDE) provide insights into the breathiness and complexity of the voice signal.
频率特征包括平均基频（MDVP:Fo）、最大频率（Fhi）和最小频率（Flo）。抖动参数捕捉基音周期的逐周期变异，如Jitter(%)和Jitter:RAP；而光泽度参数量化连续嗓音周期之间的振幅不稳定程度。其他特征如噪声谐波比（NHR）和复发周期密度熵（RPDE）则提供关于嗓音气息度和复杂度的深层信息。
4. Feature Engineering and Selection | 特征工程与特征选择
Raw voice recordings are high-dimensional and noisy. Feature engineering transforms these raw signals into meaningful numerical representations. The 22 features in the UCI dataset were extracted using the Praat acoustic analysis software and MDVP. Once extracted, feature selection helps identify which attributes contribute most to accurate classification, reducing dimensionality and computational cost.
原始嗓音录音是高维度且含噪的。特征工程将这些原始信号转化为有意义的数值表示。UCI数据集中的22个特征是通过Praat声学分析软件和MDVP工具提取的。提取完成后，特征选择有助于识别哪些属性对准确分类贡献最大，从而降低维度并减少计算成本。
Common feature selection techniques include correlation analysis, mutual information, and feature importance ranking from tree-based models. For instance, if two features such as MDVP:Fo and Jitter(%) are highly correlated, one may be redundant and can be removed without significant loss of information.
常用的特征选择技术包括相关分析、互信息法以及基于树模型的特征重要性排序。例如，如果MDVP:Fo和Jitter(%)两个特征高度相关，其中一个可能冗余，可以在不显著损失信息的情况下将其移除。
A useful mathematical framework for scoring features is the Fisher discriminant ratio, which measures the separation between two classes relative to within-class variance. A higher score indicates a more discriminative feature.
一个实用的特征评分数学框架是Fisher判别比，它衡量两个类别之间的分离程度相对于类内方差的大小。分数越高，表明特征的判别能力越强。
5. Model Selection: Decision Trees, SVM, and Neural Networks | 模型选择：决策树、支持向量机与神经网络
Three classic algorithms are frequently applied to this problem. A decision tree partitions the feature space using a series of if-then-else rules. At each node, the algorithm selects the attribute that best splits the data according to criteria such as information gain or Gini impurity. Decision trees are highly interpretable—clinicians can trace exactly how a prediction was made—but they tend to overfit unless pruned or limited in depth.
三种经典算法常被用于此类问题。决策树通过一系列“如果-那么-否则”规则来划分特征空间。在每个节点，算法根据信息增益或基尼不纯度等准则，选择最能有效分割数据的属性。决策树具有高度的可解释性——临床医生可以精确追踪预测的生成过程——但除非进行剪枝或限制深度，否则容易过拟合。
Support vector machines aim to find a hyperplane that maximises the margin between the two classes. For non-linearly separable data, the kernel trick maps inputs into a higher-dimensional space where a linear separator exists. A radial basis function (RBF) kernel is commonly used for voice data, as the relationship between acoustic features and disease status is inherently non-linear.
支持向量机的目标是找到一个使两个类别之间间隔最大化的超平面。对于非线性可分的数据，核技巧将输入映射到更高维空间中，在该空间中存在线性分隔面。径向基函数（RBF）核常被用于嗓音数据，因为声学特征与疾病状态之间的关系本质上是非线性的。
Artificial neural networks, particularly feedforward networks with hidden layers, can model complex non-linear functions. A typical architecture for this task might have an input layer of 22 neurons, one or two hidden layers with 8-16 neurons each using ReLU activation, and an output layer with sigmoid activation producing a probability between 0 and 1. The network learns weights via backpropagation and gradient descent.
人工神经网络，特别是带隐藏层的前馈网络，能够建模复杂的非线性函数。此任务的典型架构可以是：输入层22个神经元，一个或两个隐藏层每层8-16个神经元并使用ReLU激活函数，输出层使用Sigmoid激活函数生成0到1之间的概率。网络通过反向传播和梯度下降学习权重。
6. Training and Testing Splits | 训练集与测试集划分
A fundamental principle in supervised learning is that a model must be evaluated on data it has never seen during training. The dataset is typically partitioned into a training set (e.g., 80% of instances) and a test set (e.g., 20%). The model learns its parameters from the training set only, after which its generalisation ability is measured on the test set.
监督学习的一个基本原则是：模型必须在训练期间从未见过的数据上进行评估。数据集通常被划分为训练集（例如80%的样本）和测试集（例如20%的样本）。模型仅从训练集学习参数，然后通过在测试集上的表现来衡量其泛化能力。
For small datasets—the UCI Parkinson dataset contains only 195 instances—this simple split can be unstable. If the test set happens to be unrepresentative, reported accuracy might be misleading. To mitigate this, k-fold cross-validation is used: the data is divided into k folds, and the model is trained k times, each time using k−1 folds for training and the remaining fold for validation. The final performance is the average across all k validation runs.
对于小型数据集——UCI帕金森数据集仅包含195条样本——简单划分可能不稳定。如果测试集恰好不具有代表性，报告的准确率可能具有误导性。为缓解这一问题，可使用k折交叉验证：将数据划分为k折，模型训练k次，每次使用k−1折进行训练、剩余一折用于验证。最终性能为所有k次验证结果的平均值。
Formally, if the dataset is D = {d₁, d₂, …, dₙ}, then D is shuffled and split into k equal-sized subsets D₁, D₂, …, Dₖ. For iteration i, the model trains on D \ Dᵢ and validates on Dᵢ.
形式化地说，如果数据集为 D = {d₁, d₂, …, dₙ}，则将其打乱并划分为k个大小相等的子集 D₁, D₂, …, Dₖ。在第i次迭代中，模型在 D \ Dᵢ 上训练并在 Dᵢ 上验证。
7. Evaluation Metrics: Accuracy and Beyond | 评估指标：准确率及其局限
Accuracy is the most intuitive metric: it is the proportion of correctly classified instances among all instances. However, accuracy alone is insufficient when class distributions are imbalanced—for example, when 74% of the dataset belongs to the positive class, a naive classifier that always predicts ‘patient’ would achieve 74% accuracy without learning anything.
准确率是最直观的指标：它是所有样本中被正确分类样本所占的比例。然而，当类别分布不平衡时，仅靠准确率是不够的——例如，当74%的数据属于正类时，一个总是预测“患者”的朴素分类器就能达到74%的准确率，而实际上它什么也没学到。
We therefore consider the confusion matrix, a 2 × 2 table containing true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN). From this matrix we derive three key metrics:
因此我们需要引入混淆矩阵，它是一个2 × 2表格，包含真正例（TP）、真负例（TN）、假正例（FP）和假负例（FN）。由此矩阵可以推导出三个关键指标：
- Precision（精确率）= TP / (TP + FP) — the proportion of positive predictions that are actually correct.
- Recall or sensitivity（召回率/敏感度）= TP / (TP + FN) — the proportion of actual positives that are correctly identified.
- F1-score（F1分数）= 2 × (Precision × Recall) / (Precision + Recall) — the harmonic mean of precision and recall.
The overall accuracy is:
总体准确率为：
Accuracy = (TP + TN) / (TP + TN + FP + FN)
In medical diagnosis, recall is especially critical: a false negative means a sick patient is sent home untreated. Meanwhile, precision matters for avoiding unnecessary treatment and patient anxiety caused by false alarms.
在医学诊断中，召回率尤为关键：漏诊意味着一位真正的患者被当作健康人而未得到治疗。同时，精确率对于避免因误报导致不必要的治疗和患者焦虑也极为重要。
8. The ROC Curve and AUC | ROC曲线与AUC
For models that output probabilities rather than hard labels, the receiver operating characteristic (ROC) curve visualises the trade-off between true positive rate (sensitivity) and false positive rate (1 − specificity) across all possible classification thresholds. Each threshold yields a different point on the curve.
对于输出概率而非硬标签的模型，接收者操作特征（ROC）曲线在全部可能的分类阈值下，可视化真正例率（敏感度）与假正例率（1 − 特异度）之间的权衡。每个阈值对应曲线上的不同点。
The area under the ROC curve (AUC) summarises overall model quality as a single number between 0.5 (random guessing) and 1.0 (perfect discrimination). An AUC of 0.95, commonly reported for SVM models on Parkinson’s voice data, means that given a random pair of one patient and one healthy control, the model correctly ranks the patient as more likely to have the disease 95% of the time.
ROC曲线下面积（AUC）将模型的整体质量综合为一个介于0.5（随机猜测）和1.0（完美判别）之间的数值。帕金森嗓音数据上的SVM模型通常报告的AUC为0.95，这意味着随机抽取一个患者和一个健康对照，模型在95%的情况下能正确判定患者更可能患有该疾病。
9. Overfitting and Regularisation | 过拟合与正则化
Overfitting occurs when a model learns the training data too well, capturing noise rather than underlying patterns. Symptoms include high training accuracy but poor test accuracy. In the Parkinson’s context, an overfitted model might memorise the specific voice idiosyncrasies of the 23 patients in the dataset, failing to generalise to new patients.
过拟合发生在模型对训练数据学习得过好时，捕获的是噪声而非潜在模式。其症状表现为训练准确率很高但测试准确率很差。在帕金森诊断语境下，过拟合的模型可能记住了数据集中23位患者的具体嗓音特征，而无法泛化到新患者。
Several strategies combat overfitting. First, regularisation adds a penalty term to the loss function, discouraging large weights. For example, L2 regularisation modifies the loss function to include λ∑wᵢ², where λ controls the strength of penalisation. Second, decision trees can be pruned or constrained by maximum depth. Third, dropout in neural networks randomly deactivates neurons during training, forcing the network to learn more robust features.
有几种策略可以对抗过拟合。首先，正则化在损失函数中加入惩罚项，抑制过大的权重。例如，L2正则化将损失函数修改为包含 λ∑wᵢ² 的形式，其中λ控制惩罚的强度。其次，决策树可以通过剪枝或限制最大深度来约束。第三，神经网络中的Dropout在训练时随机停用神经元，迫使网络学习更稳健的特征。
10. Challenges in Real-World Deployment | 实际部署中的挑战
Despite promising laboratory results, deploying ML models for Parkinson’s diagnosis in clinical settings faces significant hurdles. Data quality is a primary concern: voice recordings collected in noisy clinic environments differ from quiet laboratory conditions, degrading model performance. The UCI dataset is also limited—195 samples from only 31 subjects mean the model may not capture the full demographic and linguistic diversity of Parkinson’s patients.
尽管实验室结果令人振奋，但在临床环境中部署用于帕金森诊断的机器学习模型仍面临重大障碍。数据质量是首要关切：在嘈杂的诊所环境中采集的嗓音录音与安静的实验室条件不同，会导致模型性能下降。UCI数据集也很有限——仅来自31位受试者的195条样本意味着模型可能无法涵盖帕金森患者完整的年龄人口统计和语言多样性。
Model interpretability poses another challenge. A clinician cannot prescribe deep-brain stimulation or dopamine replacement therapy based on a black-box neural network output alone. Explainable AI techniques such as SHAP (SHapley Additive exPlanations) values help identify which voice features most influenced a model’s prediction, providing a bridge between computational output and clinical judgement.
模型的可解释性构成另一个挑战。临床医生不能仅凭黑箱神经网络的输出就决定实施脑深部刺激或多巴胺替代疗法。可解释人工智能技术，如SHAP（SHapley加性解释）值，可以识别出哪些嗓音特征对模型的预测影响最大，从而在计算输出与临床判断之间架起桥梁。
11. Ethical Considerations and Data Privacy | 伦理考量与数据隐私
Medical machine learning carries profound ethical responsibilities. Training data often contain sensitive health information, and regulations such as the UK GDPR and the Data Protection Act 2018 impose strict requirements on data storage, anonymisation, and consent. Researchers must ensure that patient identities cannot be re-identified from de-identified datasets.
医学机器学习承担着深远的伦理责任。训练数据通常包含敏感的健康信息，英国GDPR和《2018年数据保护法》等法规对数据存储、匿名化和知情同意提出了严格要求。研究人员必须确保无法从去标识化数据集中重新识别患者身份。
Bias is equally concerning. If a model is trained predominantly on male voices, it may underperform on female patients, whose fundamental frequency is typically higher. Models must be validated on diverse populations to avoid exacerbating existing health inequalities. Furthermore, algorithmic diagnosis should serve as a decision-support tool, not a replacement for qualified medical professionals.
偏差同样令人担忧。如果模型主要基于男性嗓音训练，可能在女性患者——其基频通常更高——身上表现不佳。模型必须在多样化的人群中验证，避免加剧现有的健康不平等。此外，算法诊断应作为决策支持工具，而非取代合格医疗专业人员的角色。
12. Future Directions and AI Ethics | 未来方向与人工智能伦理
The future of ML in Parkinson’s diagnosis lies in richer data modalities and more sophisticated architectures. Wearable sensors can continuously collect movement data, enabling models to detect subtle gait abnormalities. Longitudinal studies can track disease progression, turning a binary classification problem into a regression problem that predicts UPDRS (Unified Parkinson’s Disease Rating Scale) scores over time. Deep learning models, including convolutional neural networks applied directly to raw audio spectrograms, may eliminate the need for manually engineered features.
机器学习在帕金森诊断中的未来在于更丰富的数据模态和更先进的架构。可穿戴传感器可以持续采集运动数据，使模型能够检测细微的步态异常。纵向研究可以追踪疾病进展，将二分类问题转化为预测UPDRS（统一帕金森病评定量表）评分随时间变化的回归问题。深度学习模型，包括直接应用于原始音频频谱图的卷积神经网络，可能消除对人工设计特征的需求。
For A-Level Computer Science students, this application area exemplifies the complete data science pipeline: data collection, cleaning, feature engineering, model selection, training, evaluation, and ethical reflection. It demonstrates how abstract concepts such as information gain, gradient descent, and confusion matrices translate into technologies that can potentially transform patient lives. As ML models become more accurate and interpretable, their integration into clinical workflows promises earlier detection, better treatment monitoring, and ultimately, improved quality of life for millions living with Parkinson’s disease.
对于A-Level计算机科学学生而言，这一应用领域展示了完整的数据科学流程：数据采集、清洗、特征工程、模型选择、训练、评估和伦理反思。它展示了信息增益、梯度下降和混淆矩阵等抽象概念如何转化为有可能改变患者生活的技术。随着机器学习模型变得更加准确和可解释，它们与临床工作流程的融合预示着更早的诊断、更好的治疗监测，并最终改善数百万帕金森病患者的生活质量。
Published by TutorHao | Computer Science Revision Series | aleveler.com
更多咨询请联系16621398022（同微信）
屏轩国际教育cambridge primary/secondary checkpoint, cat4, ukiset,ukcat,igcse,alevel,PAT,STEP,MAT, ibdp,ap,ssat,sat,sat2课程辅导，国外大学本科硕士研究生博士课程论文辅导