Hepatocellular carcinoma (HCC) is a highly lethal primary liver cancer, posing a challenge due to its propensity for early recurrence (ER) post-surgery. Accurate prediction of ER is pivotal for patient survival. Current research predominantly focuses on whole slide image (WSI) analysis for predicting postoperative ER of HCC, employing annotation-free and annotation-based methodologies. Annotation-free approaches, primarily using multi-instance learning (MIL), demand extensive datasets and computational resources due to the absence of pathological priors during training. Conversely, annotation-based methods, such as improved MIL, capitalize on incorporating these priors. However, the labor-intensive nature of full annotation due to the complex boundaries and large scales of WSIs remains a significant challenge. In response, we propose a novel multi-type tissue heatmap-based approach using partial annotations for ER prediction of HCC. Diverging from traditional methods reliant on exhaustive annotations, our approach significantly reduces annotation requirements while generating informative multi-type tissue heatmaps. Our partial annotations cover less than 4 \(\%\) of the total tissue area in WSIs, facilitating more efficient and cost-effective analysis. Subsequently, we employ a prediction model on these heatmaps to achieve our objective. Rigorous experimental evaluations demonstrate that utilizing tissue heatmaps of more types yields superior performance in ER prediction. Furthermore, our proposed method based on multi-type tissue heatmaps outperforms conventional MIL model and exhibits comparable performance to the improved MIL model based on ER-ProbMaps in the ER prediction task.