Reviewers' comments:

Reviewer 1: This paper did an interesting field study for a lane change prediction algorithm. This algorithm is validated in simulation first and then implemented in the real world. LSTM model is trained for each driver and used for prediction. The prediction result is analyzed and evaluated. Overall, this paper has a clear structure and methodology is well explained.
My suggestions are as below:
1. The setup for simulation is not very clear. How many simulations have been run and what scenarios are simulated is not well explained.
2. How many datasets are collected to train and validate the model in real word? Is the collected sample large enough?
3. The author might also need to give a ratio of wrong predictions and what will it cause in the real world and how to improve it.

Reviewer 2: The manuscript focuses on the mixed traffic conditions where CAVs and HDVs share the same road space. In the proposed framework for lane change prediction system is built based on the vehicle-edge-cloud architecture: specifically, driver behavior model for each connected HDV is processed on the cloud server while the edge server is used to predict the lane change maneuver. To demonstrate the framework, a field experiment was implemented and concluded that the lane change decision can be predicted 6 seconds in advance with an averaged error as 1.03 meters within a 4-second prediction window. This paper is overall well-structured and has a clear logical flow. Here are some comments and questions related to the manuscript:
1. How many data points are needed to train personalized behavior model? Since as mentioned in the abstract that the driving behavior can be affected by interactions with other vehicles, and the difficulty of scenario reproduction affect the data collection, please provide more details about the how may data are needed and how to justify the assumption that the driving behavior are consistent or can be assumed to be long-lasting for each driver?
2. Since the scenario in the paper is specifically about on/off ramp, can the personalized driving model be adapted to other situations like lane-changing maneuvers at other road segments? The reason for the question is related to the transferability of the model. Since the necessity and triggers of the lane-changing maneuvers are totally different for the two cases, one is mandatory lane change because of the road configuration while the other is discretionary lane change.
3. Does the personalized driving model consider interactions between HDVs and CAVs? The authors do mention such interaction may affect the driving behavior of human drivers, while it was not shown in the field implementation.
4. How to deal with the HDVs which are not connected? Since the ego vehicle is equipped with some perception system like camera, radar. Is it possible to construct a personalized driving model using perception data?