2.1. Data collection and pre-processing

We downloaded the Pitt dataset from the DementiaBank database (Lanzi et al., Reference Lanzi, Saylor, Fromm, Liu, MacWhinney and Cohen2023). The Pitt dataset is based on a longitudinal AD study where subjects were followed up in multiple years (Becker et al., Reference Becker, Boiler, Lopez, Saxton and McGonigle1994). Given the focus of this study on speech-based MCI detection, we included all subjects with speech data available and labeled with normal control and MCI. According to the description of the Pitt dataset (Boller and Becker, Reference Boller and Beckern.d.), normal control is defined by the diagnosis category coded 8 (800 or 821), and MCI is defined by the diagnosis category coded 6 or 7 (600, 610, 611, 720, or 740).

We extracted baseline demographics, including age, gender, race, and education (in years), from the Pitt dataset. We also extracted the baseline and follow-up cognitive scores derived from the Mini-Mental State Examination (MMSE). Further, and most importantly, we obtained the baseline and follow-up audio recordings of the Cookie Theft test, the most used picture description test in clinical settings (Eyigoz et al., Reference Eyigoz, Mathur, Santamaria, Cecchi and Naylor2020), from the Pitt dataset. We selected subjects with normal control and MCI labels, including 89 normal controls and 18 MCI subjects. Since the longitudinal observations were sparse and limited, we considered the data points cross-sectional samples to enlarge the sample size, resulting in 129 audio samples labeled with normal control and 24 with MCI. Among these 153 samples, all demographic information was complete, and the missing ratio of MMSE was 0.7%. Table 1 summarizes the descriptive statistics of the selected audio samples. As shown in Table 1, the difference in MMSE was small, suggesting that MMSE may not be a good indicator to distinguish MCI from normal control and detect the subtle changes in cognitive function in the early stages of AD.

Table 1. Data descriptive statistics

We transcribed the collected audio recordings in English to text data using OpenAI’s Whisper model (Radford et al., Reference Radford, Kim, Xu, Brockman, McLeavey and Sutskever2023), a text-to-speech model pre-trained on a large amount of audio data. In addition to the audio input, we provided a simple text prompt, “Umm, let me think like, hmm… Okay, here’s what I’m, like, thinking.,” to the model to capture the filler words in the transcript according to the official documentation for speech-to-text prompting (OpenAI, n.d.-c). We removed non-English characters in the resulting transcriptions before further analysis. Finally, we generated a tabular dataset with seven columns, including age, gender, race, education, MMSE, transcription text, and diagnosis label.

2.2. LLM-based data generation framework

After data collection and pre-processing, an LLM-based data generation framework was developed to enlarge the pre-processed text dataset, which was of small sample size and highly imbalanced (24 MCI samples out of 153 samples), by generating more text samples that are synthetic yet realistic, leveraging on the prior knowledge encoded in LLMs pre-trained on the massive amount of data. The newly generated samples were used for model development only, not model evaluation. Specifically, three data generation strategies were investigated, including (1) observation data generation, (2) cross-lingual data generation, and (3) counterfactual data generation. At a high level, observation data generation aims to generate new MCI-like text samples based on the observed MCI samples from the existing dataset. Building on top of the observational data generation strategy, two novel generation strategies were proposed. On the one hand, cross-lingual data generation further translates the new MCI-like samples into another language (Chinese in this study), introducing cross-linguistic diversity in the training data and facilitating out-of-distribution learning. On the other hand, counterfactual data generation aims to generate new MCI-like text samples based on the observed normal control samples from the existing dataset while controlling for other variables as much as possible. The counterfactual generation process requires a deeper understanding of the disease mechanism, exploiting the causal knowledge encoded in LLMs, to answer a what-if question, such as what the speech observed from the normal control subject would be if he/she was an MCI subject while holding other information unchanged.

Each LLM-based data generation strategy was implemented through prompting engineering using OpenAI’s text generation application programming interfaces (APIs) (OpenAI, n.d.-b). The prompt consists of two parts: a system message and a user message. The system message is optional and contains instructions on how the AI-driven conversation system should behave, for example, how to rephrase the text. The user message gives a specific request for the AI system to respond, for example, a text example (OpenAI, n.d.-b). In this study, the input to the OpenAI API was formatted with a two-part system message first, followed by a user message. Specifically, the first part of the system message provides background information for the data generation task.

2.3. Text classification model training and evaluation

The original and generated text samples were first vectorized before model training and evaluation. The term frequency-inverse document frequency (TF-IDF) method was adopted for text vectorization. Unlike advanced text embedding methods, which generate a high-dimensional vector for each text sample based on deep learning models such as transformers (Reimers and Gurevych, Reference Reimers and Gurevych2019), the traditional TF-IDF method was selected to obtain more explainable features to be investigated in feature importance analysis, allowing for a deeper understanding of how data generation can help alleviate biases in MCI prediction.

Irrelevant keywords were removed, and filler words were unified before TF-IDF vectorization. Specifically, irrelevant keywords included “okay,” “alright,” “tell,” “see,” “describe,” “picture,” “action,” and “happening.” These words were often observed at the beginning of the recording, asking the test subject to perform the picture description task, for example, “I want you to tell me all of the action that you see, okay?.” Punctuations and filler words, including “…,” “uh,” “um” (“umm” and “ummm”), and “hm” (“hmm” and “hmmm”), were converted to a unified representation (a single keyword “PAUSE”) to capture the pauses and hesitations in speech data. Moreover, Chinese text (from cross-lingual generation) was tokenized using Chinese text segmentation software (Sun, Reference Sunn.d.).

Taking the TF-IDF vectors as the inputs, text classification models were constructed based on eXtreme Gradient Boosting (XGBoost), a tree-boosting algorithm with 100 estimators using an objective function for binary logistic regression. MCI classification models were trained across various scenarios: (1) the original data, (2) the original data plus the three proposed data generation strategies, and (3) the original data plus different combinations of data generation strategies. For the XGBoost model trained on the original data, the parameter to control the balance of positive and negative weights for imbalanced labels was set to the ratio of the original normal control and MCI samples. This parameter was set to the default value of one for other scenarios with newly generated MCI samples to overcome the limited MCI samples in the dataset.

We performed five-fold cross-validation to evaluate the performance of MCI classification models using text data (see Supplementary Figure S1 for an overview). Each fold was a 20% subsample of the original whole dataset. Four folds were selected as the training set and the remaining one as the testing set. All newly generated samples for model training were obtained from LLM-based data generation using the original samples from the training set (the four folds) but not from the testing set (the remaining fold). This cross-validation procedure was repeated five times to ensure that all folds were used for testing. As a result, each MCI sample was tested precisely once, thus providing a better understanding of MCI classification performance, even though the number of MCI samples was limited. The same cross-validation folds were applied to each MCI classification model under different data generation strategies.

Two performance evaluation metrics, including sensitivity and F1-score, were adopted in this study. Sensitivity (also known as recall or true positive rate) measures the true positive rate. F1-score provides a more balanced view of predictive accuracy given imbalanced labels. However, the traditional accuracy metric, which measures the percentage of correctly labeled samples, was not used in this study because it can be misleading when dealing with imbalanced labels. For example, given the imbalanced distribution of positive and negative samples in this study (24 MCI and 129 normal control samples), a naive classifier that always predicts normal control can still give an 84% accuracy but fails to predict MCI every time, making it not useful for AD screening at all.

The two metrics used in this study are further elaborated in Equations (1) and (2). Average sensitivity and F1-score across the five-fold cross-validation were reported in this study. The higher the sensitivity and F1-score values, the better the predictive performance.

(1) $$ \mathrm{Sensitivity}= TP/\left( TP+ FN\right) $$

(2) $$ {\displaystyle \begin{array}{c}\mathrm{F}1-\mathrm{score}=\frac{TP}{TP+\frac{1}{2}\left( FP+ FN\right)}\end{array}} $$

where TP, FN, and FP are obtained from the confusion matrix calculated based on the predicted and ground truth labels, representing the number of true positives (MCI labels that are correctly predicted), false negatives (MCI labels that are incorrectly labeled), and false positives (normal control labels that are incorrectly labeled), respectively.

2.4. SHAP analysis for identifying the most important speech markers

After model evaluation, feature importance analysis was performed to investigate the key speech features predicting MCI compared to normal control, allowing for a better understanding of the new insights brought by LLM-based text data generation and why the biases in MCI prediction can or cannot be reduced via new data generation. The feature importance score was calculated using SHapley Additive exPlanations (SHAP) analysis (Lundberg and Lee, Reference Lundberg and Lee2017). In general, SHAP analysis uses a cooperative game theoretic approach to allocate credits for a model’s output among its input features. Here, game theory is connected to AI by matching input features with players in a game and aligning the model function with the game’s rules. A feature participates in a model when information is available about its value. The importance of each feature is quantified by its contribution to the model prediction across all possible orderings of feature coalitions (Lundberg and Lee, Reference Lundberg and Lee2017).

Three models were selected for SHAP analysis, including (1) the baseline model trained on the original data, (2) the best model trained based on one of the three data generation strategies, and (3) the best model trained based on the best combination of data generation strategies. For XGBoost models, tree-based SHAP analysis was performed (Lundberg et al., Reference Lundberg, Erion, Chen, DeGrave, Prutkin, Nair, Katz, Himmelfarb, Bansal and Lee2020). The SHAP value of each feature was calculated, utilizing both training and testing sets. Features were ordered by their SHAP values, highlighting their high impacts on individual samples using the maximum absolute SHAP value, that is, features having significant positive or negative effects on MCI onset.

3.1. Evaluation of data generation strategies

Table 2 shows the data generation statistics using different strategies and LLM models. The expected number of new MCI samples obtained from observational and cross-lingual generation is 120, given that these two strategies were repeated five times. The expected number of new MCI samples obtained from counterfactual generation is 129 (i.e., the number of normal control samples). However, a few samples were dropped, mainly due to the following reasons: (1) some generated text samples failed to use the specific format provided in the prompt, and no new transcription data could be extracted, and (2) some were automatically blocked due to the potentially sensitive content according to the content filtering of Microsoft Azure OpenAI service.

Table 2. Data generation statistics across different strategies and models ^a

^a Data generation strategy abbreviation: OBG, observational generation; CLG, cross-lingual generation; CFG, counterfactual generation. The Euclidean distance measures the average distance between the original and newly generated samples in the high-dimensional TF-IDF vector space, where a lower value indicates a higher similarity between the original and new samples.

For each data generation strategy across different LLM models (including Gemma-2, Llama-3.1, GPT-3.5, and GPT-4), we calculated two mean text vectors based on the original and newly generated samples using the TF-IDF method. The average distance was calculated using the Euclidean distance between the two mean text vectors. Results show that the average distance varied across different data generation strategies. Observational data generation, which aims to mimic patterns observed in the original samples without introducing substantial variations, has yielded the smallest distances on average, especially with GPT-4, indicating that the generated samples are more consistent with the original samples. One exception is Gemma-2, which shows a relatively higher distance compared to other models, suggesting that it may struggle to replicate the original data distribution closely. Cross-lingual data generation has yielded the largest distances likely due to the variations introduced by cross-lingual transformations. This finding is consistent across different models, suggesting that such variations are more likely to be driven by the data generation strategy rather than the model’s generation capabilities. In terms of counterfactual data generation, GPT-4 and Llama-3.1 have generated samples with higher distances from the original ones, suggesting that these models are more capable of generating complex and diverse samples in hypothetical scenarios.

Five-fold cross-validation was performed based on the original and newly generated data. The generated data were used only for model training. Table 3 lists the performance of MCI prediction using different data generation strategies, demonstrating their effects on facilitating model training and improving MCI prediction performance in terms of sensitivity and F1-score. All data generation methods have achieved higher predictive performance than the baseline model using the original data only. GPT-4 consistently delivered robust performance across all strategies, particularly excelling in observational generation while maintaining a high F1-score in counterfactual generation. Among open-source LLMs, Llama-3.1 demonstrated exceptional performance in counterfactual generation, while Gemma-2 performed less well than other models across all data generation strategies. Specifically, among LLMs employing observational generation, GPT-4 achieved the best sensitivity (24%) and F1-score (24%), outperforming GPT-3.5 and open-source models, including Gemma-2 and Llama-3.1. In terms of cross-lingual generation, performance varied across LLMs. Llama-3.1 outperformed all other models, achieving a sensitivity of 20% and the highest F1-score of 26%. Gemma-2 showed minimal improvement, while GPT models were tied at modest values of 8% sensitivity and 11% F1-score. The counterfactual generation strategy demonstrated the greatest potential for improving sensitivity. Llama-3.1 achieved the highest sensitivity (37%) and an F1-score of 30%, matching GPT-4’s F1-score, although GPT-4 trailed slightly in sensitivity (30%). GPT-3.5 and Gemma-2 performed less competitively, with GPT-3.5 reaching a sensitivity of 21% and an F1-score of 15%.

Table 3. Evaluation of different data generation strategies in improving MCI detection sensitivity and F1-score based on five-fold cross-validation ^a

^a Data generation strategy abbreviation: OBG, observational generation; CLG, cross-lingual generation; CFG, counterfactual generation. Higher sensitivity and F1-score values indicate better predictive performance. Bold value indicates the best performance.

Moreover, as shown in Table 3, counterfactual generation has demonstrated superior performance compared to other data generation methods, especially when using Llama-3.1 and GPT-4 models (which have generated counterfactual samples with higher distances from the original ones, as shown in Table 2). This advantage can be attributed to its reliance on causal reasoning, which enables the introduction of hypothetical “what-if” scenarios and reduces spurious correlations between observed features and their labels (Ilse et al., Reference Ilse, Tomczak and Forré2021). This process has been shown to effectively address dataset biases, enhancing out-of-distribution generalization performance in computer vision tasks such as imaging classification (Chang et al., Reference Chang, Adam and Goldenberg2021) and language processing tasks such as question-answering (Sachdeva et al., Reference Sachdeva, Tutek and Gurevych2023). In contrast, the cross-lingual generation has only achieved the lowest improvement on the baseline. One possible reason is that the simple TF-IDF method may only partially capture the cross-lingual text representations. Although the TF-IDF vector is a more interpretable representation, it may fail to capture the contexts and semantic relationships compared to more advanced methods, such as deep learning-based multilingual text representation learning (Cahyawijaya et al., Reference Cahyawijaya, Chen, Bang, Khalatbari, Wilie, Ji, Ishii and Fung2024). Nonetheless, as shown in Table 2, it can still provide new samples that are different from the original monolingual samples.

Furthermore, Table 4 evaluates predictive performance using different combinations of data generation strategies. These combinations generally led to improvements over single-strategy approaches, with notable variations among models. The best performance (sensitivity and F1-score) out of all settings has been achieved using all three data generation methods based on GPT-4. In contrast, for the Llama model, combining multiple strategies did not surpass the performance achieved with counterfactual generation alone. Meanwhile, the Gemma model achieved its best performance when combining cross-lingual and counterfactual generation. These findings highlight the additive effect of data generation strategies on improving predictive performance, particularly with GPT models. Compared to Gemma and Llama models, GPT models are better equipped to capture diverse aspects of speech-based MCI detection, introducing beneficial variations into the training data that enhance model generalization and robustness.

Table 4. Evaluation of different data generation combinations in improving MCI detection sensitivity and F1-score based on five-fold cross-validation ^a

^a Data generation strategy abbreviation: OBG, observational generation; CLG, cross-lingual generation; CFG, counterfactual generation. Higher sensitivity and F1-score values indicate better predictive performance. Bold value indicates the best performance.

3.2. Key speech markers distinguishing MCI from normal control

In addition to performance evaluation, SHAP analysis was performed to understand which speech markers are most important in MCI prediction before and after data generation. This feature importance analysis can facilitate understanding why LLM-based data generation can improve predictive performance and reduce biases, bringing new insights into the key speech markers predicting MCI. Figure 2 shows the top 10 speech markers predicting MCI compared to normal control based on the baseline model trained on the original data. According to Figure 2, the top maker is PAUSE. The corresponding red circles are mainly distributed across the right part of the x-axis with larger positive SHAP values. This means that filler words with higher TF-IDF values significantly contribute to MCI label prediction, suggesting that pauses are more frequently observed among the MCI transcription samples. Similarly, “reaching,” “onto,” and “dish” with higher TF-IDF values are significant contributors to MCI label prediction, making them the distinguishing markers of MCI compared to normal control.

Figure 2. The top 10 speech markers predicting MCI compared to normal control based on the baseline model trained on the original data. Each circle in the plot represents one sample. The color of the circle indicates the speech marker’s TF-IDF value (see the color bar on the right). The higher the TF-IDF value, the darker the red color. The lower the TF-IDF value, the lighter the blue color. The x-axis represents the SHAP value (i.e., the feature importance score). A higher positive value indicates a higher contribution to the prediction of the positive label (i.e., MCI). A lower negative value indicates a higher contribution to the prediction of the negative label (i.e., normal control).

Moreover, Figures 3 and 4 illustrate the top 10 speech markers predicting MCI compared to normal control based on (1) the best model using the original data and the counterfactual data generation using GPT-4 (with an F1-score of 30%; see Table 3) and (2) the best model using the original data and all data generation (i.e., observational, cross-lingual, and counterfactual generation) using GPT-4 (with an F1-score of 35%; see Table 4). Supplementary Table S1 further shows the top 50 speech markers predicting MCI compared to normal control. After incorporating the generated data, PAUSE remains the top marker in different data generation scenarios, suggesting that interruptions in speech or delays between words could be a significant indicator of MCI. MCI subjects may show more frequent pauses, potentially reflecting cognitive processing difficulties in the early stage of AD (Pistono et al., Reference Pistono, Pariente, Bézy, Lemesle, Le Men and Jucla2019). However, as shown in Figure 3, new markers, such as “something” and “might,” appear more prominently after incorporating counterfactual data generation. These words, which were insignificant in the original dataset, could reflect language deficiencies common in MCI subjects, such as uncertainty or vagueness in speech (Gosztolya et al., Reference Gosztolya, Vincze, Tóth, Pákáski, Kálmán and Hoffmann2019). The increase in the frequency of these markers in the newly generated data implies that counterfactual generation, with the help of prior knowledge encoded in LLMs when generating MCI-like samples, may help highlight subtle cognitive issues that were less significant initially. Further, the top distinctive speech markers separating the MCI samples from normal control samples have been highlighted in Figure 4, suggesting that normal control samples are more likely to use verbs, such as “running,” “falling,” and adjectives, such as “little,” compared to MCI samples, which is also evident in previous research where impaired verb fluency is considered a sign of MCI (Östberg et al., Reference Östberg, Fernaeus, Hellström, Bogdanović and Wahlund2005).

Figure 3. The top 10 speech markers predicting MCI compared to normal control based on the best model using the original data and the counterfactual data generation. Each circle in the plot represents one sample. The color of the circle indicates the speech marker’s TF-IDF value (see the color bar on the right). The higher the TF-IDF value, the darker the red color. The lower the TF-IDF value, the lighter the blue color. The x-axis represents the SHAP value (i.e., the feature importance score). A higher positive value indicates a higher contribution to the prediction of the positive label (i.e., MCI). A lower negative value indicates a higher contribution to the prediction of the negative label (i.e., normal control).

Figure 4. The top 10 speech markers predicting MCI compared to normal control based on the best model using the original data and all data generation. Each circle in the plot represents one sample. The color of the circle indicates the speech marker’s TF-IDF value (see the color bar on the right). The higher the TF-IDF value, the darker the red color. The lower the TF-IDF value, the lighter the blue color. The x-axis represents the SHAP value (i.e., the feature importance score). A higher positive value indicates a higher contribution to the prediction of the positive label (i.e., MCI). A lower negative value indicates a higher contribution to the prediction of the negative label (i.e., normal control).

3.3. Discussions

The strengths of this study are three-fold. First, unlike prior speech-based methods focused on AD dementia, this study targets early detection at the MCI stage, aiming to reduce prediction biases caused by limited and imbalanced data and improve detection sensitivity. Second, it introduces a novel LLM-based data generation framework, utilizing cross-lingual and counterfactual strategies to enhance out-of-distribution learning and improve prediction performance. Experimental results based on the Pitt corpus have shown that the proposed LLM-based data generation framework can significantly improve MCI detection sensitivity by up to 38% and F1-score by up to 31%. Third, SHAP analysis provides explainability, revealing key speech markers before and after data generation and highlighting how LLM-based data generation improves MCI prediction.

The proposed new data generation strategies, including cross-lingual and counterfactual generation, can reduce biases in MCI prediction due to the limited and imbalanced data because these two strategies can generate out-of-distribution samples either (1) with the same labels but with more linguistic variations from cross-lingual generation or (2) with opposite labels through counterfactual generation, which requires causal reasoning about the data generation process. These samples are not in the original dataset, making them helpful in model training under low-data settings. This study has particularly attempted to inject causality in LLMs using chain-of-thought prompting. The counterfactual generation has achieved the best improvement compared to other generation strategies (see Table 3), suggesting certain levels of causal reasoning capability of LLMs, which can be further verified in different scenarios.

This study, however, has several limitations that can be addressed in future work. First, the dataset for model training and evaluation is small. With the help of LLM-based data generation, the current study has demonstrated significant relative improvement (more than 30% for MCI detection sensitivity and F1-score) over the baseline, but the absolute predictive performance remains low. The trained model may remain overfit to the limited samples from the Pitt study and fail to generalize to entirely new samples, especially from a different cohort. Importantly, LLM-based data generation should not be used to replace collecting more real samples but to assist speech-based disease diagnosis in low-resource settings. Future study can incorporate more related speech datasets to improve MCI detection performance, for example, from the TAUKADIAL Challenge with MCI and normal control speech samples in English and Chinese (Luz et al., Reference Luz, Garcia, Haider, Fromm, MacWhinney, Lanzi, Chang, Chou and Liu2024). Future study can also investigate the progression modeling of AD, such as MCI-to-AD conversion, using longitudinal speech data (Xue et al., Reference Xue, Karjadi, Paschalidis, Au and Kolachalama2021). Second, some audio samples were noisy, with sounds from the examiner (though often very short and brief). The quality of audio samples may affect the performance and accuracy of speech-to-text transcription, even though this study has taken several post-hoc measures for text quality control, such as removing noisy characters that were non-English and filtering out words related to the examiner’s speech but unrelated to the picture description task itself. Future study can investigate more audio quality control techniques, for example, noise reduction and speaker diarization (Fujita et al., Reference Fujita, Kanda, Horiguchi, Xue, Nagamatsu and Watanabe2019). Third, this study utilized the text data obtained from audio transcripts without modeling audio data directly. Given that the experimental results have shown the importance of acoustic-related features, such as pauses, in predicting MCI, integrating both text and audio data using a multimodal approach can potentially improve speech-based MCI detection. Finally, while LLMs are trained on massive amounts of data, they may inadvertently encode societal biases embedded within their training datasets (Schramowski et al., Reference Schramowski, Turan, Andersen, Rothkopf and Kersting2022). These biases, such as those related to gender or ethnicity, can influence the generation of synthetic data and compromise the fairness of downstream tasks (Li et al., Reference Li, Du, Song, Wang and Wang2023). Addressing these concerns is particularly crucial in the context of real-world clinical decision-making, where biased training data can lead to inequitable outcomes (Cross et al., Reference Cross, Choma and Onofrey2024). Future study can investigate advanced LLM debiasing techniques to improve the fairness of LLM-based data generation, ultimately enhancing the reliability and equity of medical diagnostics.